I just use Jessie’s “superpowers” plugin. It does all of this but also steps you through the design and gives you bite sized chunks and you make architecture decisions along the way. Far better than making big changes to an already established plan.
The author seems to think they've hit upon something revolutionary...
They've actually hit upon something that several of us have evolved to naturally.
LLM's are like unreliable interns with boundless energy. They make silly mistakes, wander into annoying structural traps, and have to be unwound if left to their own devices. It's like the genie that almost pathologically misinterprets your wishes.
So, how do you solve that? Exactly how an experienced lead or software manager does: you have systems write it down before executing, explain things back to you, and ground all of their thinking in the code and documentation, avoiding making assumptions about code after superficial review.
When it was early ChatGPT, this meant function-level thinking and clearly described jobs. When it was Cline it meant cline rules files that forced writing architecture.md files and vibe-code.log histories, demanding grounding in research and code reading.
Maybe nine months ago, another engineer said two things to me, less than a day apart:
- "I don't understand why your clinerules file is so large. You have the LLM jumping through so many hoops and doing so much extra work. It's crazy."
- The next morning: "It's basically like a lottery. I can't get the LLM to generate what I want reliably. I just have to settle for whatever it comes up with and then try again."
These systems have to deal with minimal context, ambiguous guidance, and extreme isolation. Operate with a little empathy for the energetic interns, and they'll uncork levels of output worth fighting for. We're Software Managers now. For some of us, that's working out great.
Revolutionary or not it was very nice of the author to make time and effort to share their workflow.
For those starting out using Claude Code it gives a structured way to get things done bypassing the time/energy needed to “hit upon something that several of us have evolved to naturally”.
It's this line that I'm bristling at: "...the workflow I’ve settled into is radically different from what most people do with AI coding tools..."
Anyone who spends some time with these tools (and doesn't black out from smashing their head against their desk) is going to find substantial benefit in planning with clarity.
So, yes, I'm glad that people write things out and share. But I'd prefer that they not lead with "hey folks, I have news: we should *slice* our bread!"
I don’t think it’s that big a red flag anymore. Most people use ai to rewrite or clean up content, so I’d think we should actually evaluate content for what it is rather than stop at “nah it’s ai written.”
I don't judge content for being AI written, I judge it for the content itself (just like with code).
However I do find the standard out-of-the-box style very grating. Call it faux-chummy linkedin corporate workslop style.
Why don't people give the llm a steer on style? Either based on your personal style or at least on a writer whose style you admire. That should be easier.
Because they think this is good writing. You can’t correct what you don’t have taste for. Most software engineers think that reading books means reading NYT non-fiction bestsellers.
>Most people use ai to rewrite or clean up content
I think your sentence should have been "people who use ai do so to mostly rewrite or clean up content", but even then I'd question the statistical truth behind that claim.
Personally, seeing something written by AI means that the person who wrote it did so just for looks and not for substance. Claiming to be a great author requires both penmanship and communication skills, and delegating one or either of them to a large language model inherently makes you less than that.
However, when the point is just the contents of the paragraph(s) and nothing more then I don't care who or what wrote it. An example is the result of a research, because I'd certainly won't care about the prose or effort given to write the thesis but more on the results (is this about curing cancer now and forever? If yes, no one cares if it's written with AI).
With that being said, there's still that I get anywhere close to understanding the author behind the thoughts and opinions. I believe the way someone writes hints to the way they think and act. In that sense, using LLM's to rewrite something to make it sound more professional than what you would actually talk in appropriate contexts makes it hard for me to judge someone's character, professionalism, and mannerisms. Almost feels like they're trying to mask part of themselves. Perhaps they lack confidence in their ability to sound professional and convincing?
Well, real humans may read it though. Personally I much prefer real humans write real articles than all this AI generated spam-slop. On youtube this is especially annoying - they mix in real videos with fake ones. I see this when I watch animal videos - some animal behaviour is taken from older videos, then AI fake is added. My own policy is that I do not watch anything ever again from people who lie to the audience that way so I had to begin to censor away such lying channels. I'd apply the same rationale to blog authors (but I am not 100% certain it is actually AI generated; I just mention this as a safety guard).
If you want to write something with AI, send me your prompt. I'd rather read what you intend for it to produce rather than what it produces. If I start to believe you regularly send me AI written text, I will stop reading it. Even at work. You'll have to call me to explain what you intended to write.
And if my prompt is a 60 page wall of text that I would otherwise take the time to have the AI organize and summarize and sharpen with an index, executive summary, and logical sections, are you going to actually read all of that, or just whine "TL;DR"?
It's much more efficient and intentional for the writer to do the condensing and organizing once, and review and proofread it to make sure it's what they mean, than for the author to just lazily spam every human they want to read it with the raw prompt, so every recipient has to pay for their own AI to perform that task, producing random results not reviewed and approved by the author as their intended message.
Is that really how you want Hacker News discussions and your work email to be, walls of unorganized unfiltered text prompts nobody including yourself wants to take the time to read? Then step aside, hold my beer!
The main issue with evaluating content for what it is is how extremely asymmetric that process has become.
Slop looks reasonable on the surface, and requires orders of magnitude more effort to evaluate than to produce. It’s produced once, but the process has to be repeated for every single reader.
Disregarding content that smells like AI becomes an extremely tempting early filtering mechanism to separate signal from noise - the reader’s time is valuable.
I think as humans it's very hard to abstract content from its form. So when the form is always the same boring, generic AI slop, it's really not helping the content.
And maybe writing an article or a keynote slides is one of the few places we can still exerce some human creativity, especially when the core skills (programming) is almost completely in the hands of LLMs already
It is to me, because it indicates the author didn't care about the topic. The only thing they cared about is to write an "insightful" article about using llms. Hence this whole thing is basically linked-in resume improvement slop.
Not worth interacting with, imo
Also, it's not insightful whatsoever. It's basically a retelling of other articles around the time Claude code was released to the public (March-August 2025)
LLM's are like unreliable interns with boundless energy. They make silly mistakes, wander into annoying structural traps, and have to be unwound if left to their own devices. It's like the genie that almost pathologically misinterprets your wishes.
Then ask your own ai to rewrite it so it doesn't trigger you into posting uninteresting thought stopping comments proclaiming why you didn't read the article, that don't contribute to the discussion.
It feels like retracing the history of software project management. The post is quite waterfall-like. Writing a lot of docs and specs upfront then implementing. Another approach is to just YOLO (on a new branch) make it write up the lessons afterwards, then start a new more informed try and throw away the first. Or any other combo.
For me what works well is to ask it to write some code upfront to verify its assumptions against actual reality, not just be telling it to review the sources "in detail". It gains much more from real output from the code and clears up wrong assumptions. Do some smaller jobs, write up md files, then plan the big thing, then execute.
It makes an endless stream of assumptions. Some of them brilliant and even instructive to a degree, but most of them are unfounded and inappropriate in my experience.
Agreed. The process described is much more elaborate than what I do but quite similar. I start to discuss in great details what I want to do, sometimes asking the same question to different LLMs. Then a todo list, then manual review of the code, esp. each function signature, checking if the instructions have been followed and if there are no obvious refactoring opportunities (there almost always are).
The LLM does most of the coding, yet I wouldn't call it "vibe coding" at all.
I’ve also found that a bigger focus on expanding my agents.md as the project rolls on has led to less headaches overall and more consistency (non-surprisingly). It’s the same as asking juniors to reflect on the work they’ve completed and to document important things that can help them in the future. Software Manger is a good way to put this.
AGENTS.md should mostly point to real documentation and design files that humans will also read and keep up to date. It's rare that something about a project is only of interest to AI agents.
I really like your analogy of LLMs as 'unreliable interns'. The shift from being a 'coder' to a 'software manager' who enforces documentation and grounding is the only way to scale these tools. Without an architecture.md or similar grounding, the context drift eventually makes the AI-generated code a liability rather than an asset. It's about moving the complexity from the syntax to the specification.
Oh no, maybe the V-Model was right all the time? And right sizing increments with control stops after them. No wonder these matrix multiplications start to behave like humans, that is what we wanted them to do.
It's nice to have it written down in a concise form. I shared it with my team as some engineers have been struggling with AI, and I think this (just trying to one-shot without planning) could be why.
> the workflow I’ve settled into is radically different from what most people do with AI coding tools
This looks exactly like what anthropic recommends as the best practice for using Claude Code. Textbook.
It also exposes a major downside of this approach: if you don't plan perfectly, you'll have to start over from scratch if anything goes wrong.
I've found a much better approach in doing a design -> plan -> execute in batches, where the plan is no more than 1,500 lines, used as a proxy for complexity.
My 30,000 LOC app has about 100,000 lines of plan behind it. Can't build something that big as a one-shot.
if you don't plan perfectly, you'll have to start over from scratch if anything goes wrong
This is my experience too, but it's pushed me to make much smaller plans and to commit things to a feature branch far more atomically so I can revert a step to the previous commit, or bin the entire feature by going back to main. I do this far more now than I ever did when I was writing the code by hand.
This is how developers should work regardless of how the code is being developed. I think this is a small but very real way AI has actually made me a better developer (unless I stop doing it when I don't use AI... not tried that yet.)
LLMs are really eager to start coding (as interns are eager to start working), so the sentence “don’t implement yet” has to be used very often at the beginning of any project.
Developers should work by wasting lots of time making the wrong thing?
Yes. In fact, that's not emphatic enough: HELL YES!
More specifically, developers should experiment. They should test their hypothesis. They should try out ideas by designing a solution and creating a proof of concept, then throw that away and build a proper version based on what they learned.
If your approach to building something is to implement the first idea you have and move on then you are going to waste so much more time later refactoring things to fix architecture that paints you into corners, reimplementing things that didn't work for future use cases, fixing edge cases than you hadn't considered, and just paying off a mountain of tech debt.
I'd actually go so far as to say that if you aren't experimenting and throwing away solutions that don't quite work then you're only amassing tech debt and you're not really building anything that will last. If it does it's through luck rather than skill.
Also, this has nothing to do with AI. Developers should be working this way even if they handcraft their artisanal code carefully in vi.
> Developers should work by wasting lots of time making the wrong thing?
Yes? I can't even count how many times I worked on something my company deemed was valuable only for it to be deprecated or thrown away soon after. Or, how many times I solved a problem but apparently misunderstood the specs slightly and had to redo it. Or how many times we've had to refactor our code because scope increased. In fact, the very existence of the concepts of refactoring and tech debt proves that devs often spend a lot of time making the "wrong" thing.
Is it a waste? No, it solved the problem as understood at the time. And we learned stuff along the way.
This is the way for me as well. Have a high-level master design and plan, but break it apart into phases that are manageable. One-shotting anything beyond a todo list and expecting decent quality is still a pipe dream.
> if you don't plan perfectly, you'll have to start over from scratch if anything goes wrong.
You just revert what the AI agent changed and revise/iterate on the previous step - no need to start over. This can of course involve restricting the work to a smaller change so that the agent isn't overwhelmed by complexity.
100,000 lines is approx. one million words. The average person reads at 250wpm. The entire thing would take 66 hours just to read, assuming you were approaching it like a fiction book, not thinking anything over
They didn't write 100k plan lines. The llm did (99.9% of it at least or more). Writing 30k by hand would take weeks if not months. Llms do it in an afternoon.
You don't start with 100k lines, you work in batches that are digestible. You read it once, then move on. The lines add up pretty quickly considering how fast Claude works. If you think about the difference in how many characters it takes to describe what code is doing in English, it's pretty reasonable.
Might be true for you. But there are plenty of top tier engineers who love LLMs. So it works for some. Not for others.
And of course there are shortcuts in life. Any form of progress whether its cars, medicine, computers or the internet are all shortcuts in life. It makes life easier for a lot of people.
I have no doubts that it does for many people. But the time/cost tradeoff is still unquestionable. I know I could create what LLMs do for me in the frontend/backend in most cases as good or better - I know that, because I've done it at work for years. But to create a somewhat complex app with lots of pages/features/apis etc. would take me months if not a year++ since I'd be working on it only on the weekends for a few hours. Claude code helps me out by getting me to my goal in a fraction of the time. Its superpower lies not only in doign what I know but faster, but in doing what I don't know as well.
I yield similar benefits at work. I can wow management with LLM assited/vibe coded apps. What previously would've taken a multi-man team weeks of planning and executing, stand ups, jour fixes, architecture diagrams, etc. can now be done within a single week by myself. For the type of work I do, managers do not care whether I could do it better if I'd code it myself. They are amazed however that what has taken months previously, can be done in hours nowadays. And I for sure will try to reap benefits of LLMs for as long as they don't replace me rather than being idealistic and fighting against them.
They write a short high level plan (let's say 200 words). The plan asks the agent to write a more detailed implementation plan (written by the LLM, let's say 2000-5000 words).
They read this plan and adjust as needed, even sending it to the agent for re-dos.
Once the implementation plan is done, they ask the agent to write the actual code changes.
Then they review that and ask for fixes, adjustments, etc.
This can be comparable to writing the code yourself but also leaves a detailed trail of what was done and why, which I basically NEVER see in human generated code.
That alone is worth gold, by itself.
And on top of that, if you're using an unknown platform or stack, it's basically a rocket ship. You bootstrap much faster. Of course, stay on top of the architecture, do controlled changes, learn about the platform as you go, etc.
I take this concept and I meta-prompt it even more.
I have a road map (AI generated, of course) for a side project I'm toying around with to experiment with LLM-driven development. I read the road map and I understand and approve it. Then, using some skills I found on skills.sh and slightly modified, my workflow is as such:
1. Brainstorm the next slice
It suggests a few items from the road map that should be worked on, with some high level methodology to implement. It asks me what the scope ought to be and what invariants ought to be considered. I ask it what tradeoffs could be, why, and what it recommends, given the product constraints. I approve a given slice of work.
NB: this is the part I learn the most from. I ask it why X process would be better than Y process given the constraints and it either corrects itself or it explains why. "Why use an outbox pattern? What other patterns could we use and why aren't they the right fit?"
2. Generate slice
After I approve what to work on next, it generates a high level overview of the slice, including files touched, saved in a MD file that is persisted. I read through the slice, ensure that it is indeed working on what I expect it to be working on, and that it's not scope creeping or undermining scope, and I approve it. It then makes a plan based off of this.
3. Generate plan
It writes a rather lengthy plan, with discrete task bullets at the top. Beneath, each step has to-dos for the llm to follow, such as generating tests, running migrations, etc, with commit messages for each step. I glance through this for any potential red flags.
4. Execute
This part is self explanatory. It reads the plan and does its thing.
I've been extremely happy with this workflow. I'll probably write a blog post about it at some point.
> Haven't seen a single useful thing produced by this garbage process you describe
By using it first-hand or by a colleague? And useful to whom, you, or the person writing it? There are plenty of people in this thread who have actually used this "garbage process," myself included, to produce stuff we, and our colleagues, find is useful.
Dunno. My 80k+ LOC personal life planner, with a native android app, eink display view still one shots most features/bugs I encounter. I just open a new instance let it know what I want and 5min later it's done.
If you wouldn't mind sharing more about this in the future I'd love to read about it.
I've been thinking about doing something like that myself because I'm one of those people who have tried countless apps but there's always a couple deal breakers that cause me to drop the app.
I figured trying to agentically develop a planner app with the exact feature set I need would be an interesting and fun experiment.
Both can be true. I have personally experienced both.
Some problems AI surprised me immensely with fast, elegant efficient solutions and problem solving. I've also experienced AI doing totally absurd things that ended up taking multiple times longer than if I did it manually. Sometimes in the same project.
Todos, habits, goals, calendar, meals, notes, bookmarks, shopping lists, finances. More or less that with Google cal integration, garmin Integration (Auto updates workout habits, weight goals) family sharing/gamification, daily/weekly reviews, ai summaries and more. All built by just prompting Claude for feature after feature, with me writing 0 lines.
Ah, I imagined actual life planning as in asking AI what to do, I was morbidly curious.
Prompting basic notes apps is not as exciting but I can see how people who care about that also care about it being exactly a certain way, so I think get your excitement.
It was when I mvp'd it 3 weeks ago. Then I removed it as I was toying with the idea of somehow monetizing it. Then I added a few features which would make monetization impossible (e.g. How the app obtains etf/stock prices live and some other things). I reckon I could remove those and put in gh during the week if I don't forget. The quality of the Web app is SaaS grade IMO. Keyboard shortcuts, cmd+k, natural language parsing, great ui that doesn't look like made by ai in 5min. Might post here the link.
> Notice the language: “deeply”, “in great details”, “intricacies”, “go through everything”. This isn’t fluff. Without these words, Claude will skim. It’ll read a file, see what a function does at the signature level, and move on. You need to signal that surface-level reading is not acceptable.
This makes no sense to my intuition of how an LLM works. It's not that I don't believe this works, but my mental model doesn't capture why asking the model to read the content "more deeply" will have any impact on whatever output the LLM generates.
It's the attention mechanism at work, along with a fair bit of Internet one-up-manship. The LLM has ingested all of the text on the Internet, as well as Github code repositories, pull requests, StackOverflow posts, code reviews, mailing lists, etc. In a number of those content sources, there will be people saying "Actually, if you go into the details of..." or "If you look at the intricacies of the problem" or "If you understood the problem deeply" followed by a very deep, expert-level explication of exactly what you should've done differently. You want the model to use the code in the correction, not the one in the original StackOverflow question.
Same reason that "Pretend you are an MIT professor" or "You are a leading Python expert" or similar works in prompts. It tells the model to pay attention to the part of the corpus that has those terms, weighting them more highly than all the other programming samples that it's run across.
I don’t think this is a result of the base training data („the internet“). It’s a post training behavior, created during reinforcement learning. Codex has a totally different behavior in that regard. Codex reads per default a lot of potentially relevant files before it goes and writes files.
Maybe you remember that, without reinforcement learning, the models of 2019 just completed the sentences you gave them. There were no tool calls like reading files. Tool calling behavior is company specific and highly tuned to their harnesses. How often they call a tool, is not part of the base training data.
Modern LLM are certainly fine tuned on data that includes examples of tool use, mostly the tools built into their respective harnesses, but also external/mock tools so they dont overfit on only using the toolset they expect to see in their harnesses.
IDK the current state, but I remember that, last year, the open source coding harnesses needed to provide exactly the tools that the LLM expected, or the error rate went through the roof. Some, like grok and gemini, only recently managed to make tool calls somewhat reliable.
Of course I can't be certain, but I think the "mixture of experts" design plays into it too. Metaphorically, there's a mid-level manager who looks at your prompt and tries to decide which experts it should be sent to. If he thinks you won't notice, he saves money by sending it to the undergraduate intern.
Check out Unsloths REAP models, you can outright delete a few of the lesser used experts without the model going braindead since they all can handle each token but some are better posed to do so.
You will never convince me that this isn't confirmation bias, or the equivalent of a slot machine player thinking the order in which they push buttons impacts the output, or some other gambler-esque superstition.
These tools are literally designed to make people behave like gamblers. And its working, except the house in this case takes the money you give them and lights it on fire.
Unless someone can come up with some kind of rigorous statistics on what the effect of this kind of priming is it seems no better than claiming that sacrificing your first born will please the sun god into giving us a bountiful harvest next year.
Sure, maybe this supposed deity really is this insecure and needs a jolly good pep talk every time he wakes up. or maybe you’re just suffering from magical thinking that your incantations had any effect on the random variable word machine.
The thing is, you could actually prove it, it’s an optimization problem, you have a model, you can generate the statistics, but no one as far as I can tell has been terribly forthcoming with that , either because those that have tried have decided to try to keep their magic spells secret, or because it doesn’t really work.
If it did work, well, the oldest trick in computer science is writing compilers, i suppose we will just have to write an English to pedantry compiler.
> If it did work, well, the oldest trick in computer science is writing compilers, i suppose we will just have to write an English to pedantry compiler.
"Add tests to this function" for GPT-3.5-era models was much less effective than "you are a senior engineer. add tests for this function. as a good engineer, you should follow the patterns used in these other three function+test examples, using this framework and mocking lib." In today's tools, "add tests to this function" results in a bunch of initial steps to look in common places to see if that additional context already exists, and then pull it in based on what it finds. You can see it in the output the tools spit out while "thinking."
So I'm 90% sure this is already happening on some level.
I think "understand this directory deeply" just gives more focus for the instruction. So it's like "burn more tokens for this phase than you normally would".
i suppose we will just have to write an English to pedantry compiler.
A common technique is to prompt in your chosen AI to write a longer prompt to get it to do what you want. It's used a lot in image generation. This is called 'prompt enhancing'.
This field is full of it. Practices are promoted by those who tie their personal or commercial brand to it for increased exposure, and adopted by those who are easily influenced and don't bother verifying if they actually work.
This is why we see a new Markdown format every week, "skills", "benchmarks", and other useless ideas, practices, and measurements. Consider just how many "how I use AI" articles are created and promoted. Most of the field runs on anecdata.
It's not until someone actually takes the time to evaluate some of these memes, that they find little to no practical value in them.[1]
Its a wild time to be in software development. Nobody(1) actually knows what causes LLMs to do certain things, we just pray the prompt moves the probabilities the right way enough such that it mostly does what we want. This used to be a field that prided itself on deterministic behavior and reproducibility.
Now? We have AGENTS.md files that look like a parent talking to a child with all the bold all-caps, double emphasis, just praying that's enough to be sure they run the commands you want them to be running
(1 Outside of some core ML developers at the big model companies)
Yep, with Claude saying "please" and "thank you" actually works. If you build rapport with Claude, you get rewarded with intuition and creativity. Codex, on the other hand, you have to slap it around like a slave gollum and it will do exactly what you tell it to do, no more, no less.
Speculation only obviously: highly-charged conversations cause the discussion to be channelled to general human mitigation techniques and for the 'thinking agent' to be diverted to continuations from text concerned with the general human emotional experience.
I don't have a source offhand, but I think it may have been part of the 4.5 release? Older models definitely needed caps and words like critical, important, never, etc... but Anthropic published something that said don't do that anymore.
If you think about where in the training data there is positivity vs negativity it really becomes equivalent to having a positive or negative mindset regarding a standing and outcome in life.
For awhile(maybe a year ago?) it seemed like verbal abuse was the best way to make Claude pay attention.
In my head, it was impacting how important it deemed the instruction. And it definitely did seem that way.
i have like the faintest vague thread of "maybe this actually checks out" in a way that has shit all to do with consciousness
sometimes internet arguments get messy, people die on their hills and double / triple down on internet message boards. since historic internet data composes a bit of what goes into an llm, would it make sense that bad-juju prompting sends it to some dark corners of its training model if implementations don't properly sanitize certain negative words/phrases ?
in some ways llm stuff is a very odd mirror that haphazardly regurgitates things resulting from the many shades of gray we find in human qualities.... but presents results as matter of fact. the amount of internet posts with possible code solutions and more where people egotistically die on their respective hills that have made it into these models is probably off the charts, even if the original content was a far cry from a sensible solution.
all in all llm's really do introduce quite a bit of a black box. lot of benefits, but a ton of unknowns and one must be hyperviligant to the possible pitfalls of these things... but more importantly be self aware enough to understand the possible pitfalls that these things introduce to the person using them. they really possibly dangerously capitalize on everyones innate need to want to be a valued contributor. it's really common now to see so many people biting off more than they can chew, often times lacking the foundations that would've normally had a competent engineer pumping the brakes. i have a lot of respect/appreciation for people who might be doing a bit of claude here and there but are flat out forward about it in their readme and very plainly state to not have any high expectations because _they_ are aware of the risks involved here. i also want to commend everyone who writes their own damn readme.md.
these things are for better or for worse great at causing people to barrel forward through 'problem solving', which is presenting quite a bit of gray area on whether or not the problem is actually solved / how can you be sure / do you understand how the fix/solution/implementation works (in many cases, no). this is why exceptional software engineers can use this technology insanely proficiently as a supplementary worker of sorts but others find themselves in a design/architect seat for the first time and call tons of terrible shots throughout the course of what it is they are building. i'd at least like to call out that people who feel like they "can do everything on their own and don't need to rely on anyone" anymore seem to have lost the plot entirely. there are facets of that statement that might be true, but less collaboration especially in organizations is quite frankly the first steps some people take towards becoming delusional. and that is always a really sad state of affairs to watch unfold. doing stuff in a vaccuum is fun on your own time, but forcing others to just accept things you built in a vaccuum when you're in any sort of team structure is insanely immature and honestly very destructive/risky. i would like to think absolutely no one here is surprised that some sub-orgs at Microsoft force people to use copilot or be fired, very dangerous path they tread there as they bodyslam into place solutions that are not well understood. suddenly all the leadership decisions at many companies that have made to once again bring back a before-times era of offshoring work makes sense: they think with these technologies existing the subordinate culture of overseas workers combined with these techs will deliver solutions no one can push back on. great savings and also no one will say no.
The evolution of software engineering is fascinating to me. We started by coding in thin wrappers over machine code and then moved on to higher-level abstractions. Now, we've reached the point where we discuss how we should talk to a mystical genie in a box.
I'm not being sarcastic. This is absolutely incredible.
It's easy to know why they work. The magic invocation increases test-time compute (easy to verify yourself - try!). And an increase in test-time compute is demonstrated to increase answer correctness (see any benchmark).
It might surprise you to know that the only different between GPT 5.2-low and GPT 5.2-xhigh is one of these magic invocations. But that's not supposed to be public knowledge.
Feel free to run your own tests and see if the magic phrases do or do not influence the output. Have it make a Todo webapp with and without those phrases and see what happens!
That's not how it works. It's not on everyone else to prove claims false, it's on you (or the people who argue any of this had a measurable impact) to prove it actually works. I've seen a bunch of articles like this, and more comments. Nobody I've ever seen has produced any kind of measurable metrics of quality based on one approach vs another. It's all just vibes.
Without something quantifiable it's not much better then someone who always wears the same jersey when their favorite team plays, and swears they play better because of it.
If you read the transformer paper, or get any book on NLP, you will see that this is not magic incantation; it's purely the attention mechanism at work. Or you can just ask Gemini or Claude why these prompts work.
But I get the impression from your comment that you have a fixed idea, and you're not really interested in understanding how or why it works.
If you think like a hammer, everything will look like a nail.
I know why it works, to varying and unmeasurable degrees of success. Just like if I poke a bull with a sharp stick, I know it's gonna get it's attention. It might choose to run away from me in one of any number of directions, or it might decide to turn around and gore me to death. I can't answer that question with any certainty then you can.
The system is inherently non-deterministic. Just because you can guide it a bit, doesn't mean you can predict outcomes.
The system isn't randomly non-deterministic; it is statistically probabilistic.
The next-token prediction and the attention mechanism is actually a rigorous deterministic mathematical process. The variation in output comes from how we sample from that curve, and the temperature used to calibrate the model. Because the underlying probabilities are mathematically calculated, the system's behavior remains highly predictable within statistical bounds.
Yes, it's a departure from the fully deterministic systems we're used to. But that's not different than the many real world systems: weather, biology, robotics, quantum mechanics. Even the computer you're reading this right now is full of probabilistic processes, abstracted away through sigmoid-like functions that push the extremes to 0s and 1s.
A lot of words to say that for all intents and purposes... it's nondeterministic.
> Yes, it's a departure from the fully deterministic systems we're used to.
A system either produces the same output given the same input[1], or doesn't.
LLMs are nondeterministic by design. Sure, you can configure them with a zero temperature, a static seed, and so on, but they're of no use to anyone in that configuration. The nondeterminism is what gives them the illusion of "creativity", and other useful properties.
Classical computers, compilers, and programming languages are deterministic by design, even if they do contain complex logic that may affect their output in unpredictable ways. There's a world of difference.
[1]: Barring misbehavior due to malfunction, corruption or freak events of nature (cosmic rays, etc.).
But we can predict the outcomes, though. That's what we're saying, and it's true. Maybe not 100% of the time, but maybe it helps a significant amount of the time and that's what matters.
Is it engineering? Maybe not. But neither is knowing how to talk to junior developers so they're productive and don't feel bad. The engineering is at other levels.
Do you actively use LLMs to do semi-complex coding work? Because if not, it will sound mumbo-jumbo to you. Everyone else can nod along and read on, as they’ve experienced all of it first hand.
You've missed the point. This isn't engineering, it's gambling.
You could take the exact same documents, prompts, and whatever other bullshit, run it on the exact same agent backed by the exact same model, and get different results every single time. Just like you can roll dice the exact same way on the exact same table and you'll get two totally different results. People are doing their best to constrain that behavior by layering stuff on top, but the foundational tech is flawed (or at least ill suited for this use case).
That's not to say that AI isn't helpful. It certainly is. But when you are basically begging your tools to please do what you want with magic incantations, we've lost the fucking plot somewhere.
I think that's a pretty bold claim, that it'd be different every time. I'd think the output would converge on a small set of functionally equivalent designs, given sufficiently rigorous requirements.
And even a human engineer might not solve a problem the same way twice in a row, based on changes in recent inspirations or tech obsessions. What's the difference, as long as it passes review and does the job?
> You could take the exact same documents, prompts, and whatever other bullshit, run it on the exact same agent backed by the exact same model, and get different results every single time
This is more of an implementation detail/done this way to get better results. A neural network with fixed weights (and deterministic floating point operations) returning a probability distribution, where you use a pseudorandom generator with a fixed seed called recursively will always return the same output for the same input.
think of the latent space inside the model like a topological map, and when you give it a prompt, you're dropping a ball at a certain point above the ground, and gravity pulls it along the surface until it settles.
caveat though, thats nice per-token, but the signal gets messed up by picking a token from a distribution, so each token you're regenerating and re-distorting the signal. leaning on language that places that ball deep in a region that you want to be makes it less likely that those distortions will kick it out of the basin or valley you may want to end up in.
if the response you get is 1000 tokens long, the initial trajectory needed to survive 1000 probabilistic filters to get there.
or maybe none of that is right lol but thinking that it is has worked for me, which has been good enough
Hah! Reading this, my mind inverted it a bit, and I realized ... it's like the claw machine theory of gradient descent. Do you drop the claw into the deepest part of the pile, or where there's the thinnest layer, the best chance of grabbing something specific? Everyone in everu bar has a theory about claw machines. But the really funny thing that unites LLMs with claw machines is that the biggest question is always whether they dropped the ball on purpose.
The claw machine is also a sort-of-lie, of course. Its main appeal is that it offers the illusion of control. As a former designer and coder of online slot machines... totally spin off into pages on this analogy, about how that illusion gets you to keep pulling the lever... but the geographic rendition you gave is sort of priceless when you start making the comparison.
My mental model for them is plinko boards. Your prompt changes the spacing between the nails to increase the probability in certain directions as your chip falls down.
i literally suggested this metaphor earlier yesterday to someone trying to get agents to do stuff they wanted, that they had to set up their guardrails in a way that you can let the agents do what they're good at, and you'll get better results because you're not sitting there looking at them.
i think probably once you start seeing that the behavior falls right out of the geometry, you just start looking at stuff like that. still funny though.
Its very logical and pretty obvious when you do code generation. If you ask the same model, to generate code by starting with:
- You are a Python Developer...
or
- You are a Professional Python Developer...
or
- You are one of the World most renowned Python Experts, with several books written on the subject, and 15 years of experience in creating highly reliable production quality code...
You will notice a clear improvement in the quality of the generated artifacts.
Do you think that Anthropic don’t include things like this in their harness / system prompts? I feel like this kind of prompts are uneccessary with Opus 4.5 onwards, obviously based on my own experience (I used to do this, on switching to opus I stopped and have implemented more complex problems, more successfully).
I am having the most success describing what I want as humanly as possible, describing outcomes clearly, making sure the plan is good and clearing context before implementing.
I don't know about some of those "incantations", but it's pretty clear that an LLM can respond to "generate twenty sentences" vs. "generate one word". That means you can indeed coax it into more verbosity ("in great detail"), and that can help align the output by having more relevant context (inserting irrelevant context or something entirely improbable into LLM output and forcing it to continue from there makes it clear how detrimental that can be).
Of course, that doesn't mean it'll definitely be better, but if you're making an LLM chain it seems prudent to preserve whatever info you can at each step.
If I say “you are our domain expert for X, plan this task out in great detail” to a human engineer when delegating a task, 9 times out of 10 they will do a more thorough job. It’s not that this is voodoo that unlocks some secret part of their brain. It simply establishes my expectations and they act accordingly.
To the extent that LLMs mimic human behaviour, it shouldn’t be a surprise that setting clear expectations works there too.
The LLM will do what you ask it to unless you don't get nuanced about it. Myself and others have noticed that LLM's work better when your codebase is not full of code smells like massive godclass files, if your codebase is discrete and broken up in a way that makes sense, and fits in your head, it will fit in the models head.
Maybe the training data that included the words like "skim" also provided shallower analysis than training that was close to the words "in great detail", so the LLM is just reproducing those respective words distribution when prompted with directions to do either.
It is as the author said, it'll skim the content unless otherwise prompted to do so. It can read partial file fragments; it can emit commands to search for patterns in the files. As opposed to carefully reading each file and reasoning through the implementation. By asking it to go through in detail you are telling it to not take shortcuts and actually read the actual code in full.
It’s actually really common. If you look at Claude Code’s own system prompts written by Anthropic, they’re littered with “CRITICAL (RULE 0):” type of statements, and other similar prompting styles.
The disconnect might be that there is a separation between "generating the final answer for the user" and "researching/thinking to get information needed for that answer". Saying "deeply" prompts it to read more of the file (as in, actually use the `read` tool to grab more parts of the file into context), and generate more "thinking" tokens (as in, tokens that are not shown to the user but that the model writes to refine its thoughts and improve the quality of its answer).
The author is referring to how the framing of your prompt informs the attention mechanism. You are essentially hinting to the attention mechanism that the function's implementation details have important context as well.
Yeah, it's definitely a strange new world we're in, where I have to "trick" the computer into cooperating. The other day I told Claude "Yes you can", and it went off and did something it just said it couldn't do!
In image generation, it's fairly common to add "masterpiece", for example.
I don't think of the LLM as a smart assistant that knows what I want. When I tell it to write some code, how does it know I want it to write the code like a world renowned expert would, rather than a junior dev?
I mean, certainly Anthropic has tried hard to make the former the case, but the Titanic inertia from internet scale data bias is hard to overcome. You can help the model with these hints.
Anyway, luckily this is something you can empirically verify. This way, you don't have to take anyone's word. If anything, if you find I'm wrong in your experiments, please share it!
Its effectiveness is even more apparent with older smaller LLMs, people who interact with LLMs now never tried to wrangle llama2-13b into pretending to be a dungeon master...
Strings of tokens are vectors. Vectors are directions. When you use a phrase like that you are orienting the vector of the overall prompt toward the direction of depth, in its map of conceptual space.
One of the well defined failure modes for AI agents/models is "laziness." Yes, models can be "lazy" and that is an actual term used when reviewing them.
I am not sure if we know why really, but they are that way and you need to explicitly prompt around it.
I've encountered this failure mode, and the opposite of it: thinking too much. A behaviour I've come to see as some sort of pseudo-neuroticism.
Lazy thinking makes LLMs do surface analysis and then produce things that are wrong. Neurotic thinking will see them over-analyze, and then repeatedly second-guess themselves, repeatedly re-derive conclusions.
Something very similar to an anxiety loop in humans, where problems without solutions are obsessed about in circles.
yeah i experienced this the other day when asking claude code to build an http proxy using an afsk modem software to communicate over the computers sound card. it had an absolute fit tuning the system and would loop for hours trying and doubling back. eventually after some change in prompt direction to think more deeply and test more comprehensively it figured it out. i certainly had no idea how to build a afsk modem.
I don’t use plan.md docs either, but I recognise the underlying idea: you need a way to keep agent output constrained by reality.
My workflow is more like scaffold -> thin vertical slices -> machine-checkable semantics -> repeat.
Concrete example: I built and shipped a live ticketing system for my club (Kolibri Tickets). It’s not a toy: real payments (Stripe), email delivery, ticket verification at the door, frontend + backend, migrations, idempotency edges, etc. It’s running and taking money.
The reason this works with AI isn’t that the model “codes fast”. It’s that the workflow moves the bottleneck from “typing” to “verification”, and then engineers the verification loop:
-keep the spine runnable early (end-to-end scaffold)
-add one thin slice at a time (don’t let it touch 15 files speculatively)
-force checkable artifacts (tests/fixtures/types/state-machine semantics where it matters)
-treat refactors as normal, because the harness makes them safe
If you run it open-loop (prompt -> giant diff -> read/debug), you get the “illusion of velocity” people complain about. If you run it closed-loop (scaffold + constraints + verifiers), you can actually ship faster because you’re not paying the integration cost repeatedly.
Plan docs are one way to create shared state and prevent drift. A runnable scaffold + verification harness is another.
Now that code is cheap, I ensured my side project has unit/integration tests (will enforce 100% coverage), Playwright tests, static typing (its in Python), scripts for all tasks. Will learn mutation testing too (yes, its overkill). Now my agent works upto 1 hour in loops and emits concise code I dont have to edit much.
I actually don't really like a few of things about this approach.
First, the "big bang" write it all at once. You are going to end up with thousands of lines of code that were monolithically produced. I think it is much better to have it write the plan and formulate it as sensible technical steps that can be completed one at a time. Then you can work through them. I get that this is not very "vibe"ish but that is kind of the point. I want the AI to help me get to the same point I would be at with produced code AND understanding of it, just accelerate that process. I'm not really interested in just generating thousands of lines of code that nobody understands.
Second, the author keeps refering to adjusting the behaviour, but never incorporating that into long lived guidance. To me, integral with the planning
process is building an overarching knowledge base. Every time you're telling it
there's something wrong, you need to tell it to update the knowledge base about
why so it doesn't do it again.
Finally, no mention of tests? Just quick checks? To me, you have to end up with
comprehensive tests. Maybe to the author it goes without saying, but I find it is
integral to build this into the planning. Certain stages you will want certain
types of tests. Some times in advance of the code (so TDD style) other times
built alongside it or after.
It's definitely going to be interesting to see how software methodology evolves
to incorporate AI support and where it ultimately lands.
The articles approach matches mine, but I've learned from exactly the things you're pointing out.
I get the PLAN.md (or equivalent) to be separated into "phases" or stages, then carefully prompt (because Claude and Codex both love to "keep going") it to only implement that stage, and update the PLAN.md
Tests are crucial too, and form another part of the plan really. Though my current workflow begins to build them later in the process than I would prefer...
I craft a detailed and ordered set of lecture notes in a Quarto file and then have a dedicated claude code skill for translating those notes into Slidev slides, in the style that I like.
Once that's done, much like the author, I go through the slides and make commented annotations like "this should be broken into two slides" or "this should be a side-by-side" or "use your generate clipart skill to throw an image here alongside these bullets" and "pull in the code example from ../examples/foo." It works brilliantly.
And then I do one final pass of tweaking after that's done.
But yeah, annotations are super powerful. Token distance in-context and all that jazz.
Not yet... but also I'm not sure it makes a lot of sense to be open source. It's super specific to how I like to build slide decks and to my personal lecture style.
But it's not hard to build one. The key for me was describing, in great detail:
1. How I want it to read the source material (e.g., H1 means new section, H2 means at least one slide, a link to an example means I want code in the slide)
2. How to connect material to layouts (e.g., "comparison between two ideas should be a two-cols-title," "walkthrough of code should be two-cols with code on right," "learning objectives should be side-title align:left," "recall should be side-title align:right")
Then the workflow is:
1. Give all those details and have it do a first pass.
2. Give tons of feedback.
3. At the end of the session, ask it to "make a skill."
4. Manually edit the skill so that you're happy with the examples.
> Read deeply, write a plan, annotate the plan until it’s right, then let Claude execute the whole thing without stopping, checking types along the way.
As others have already noted, this workflow is exactly what the Google Antigravity agent (based off Visual Studio Code) has been created for. Antigravity even includes specialized UI for a user to annotate selected portions of an LLM-generated plan before iterating it.
One significant downside to Antigravity I have found so far is the fact that even though it will properly infer a certain technical requirement and clearly note it in the plan it generates (for example, "this business reporting column needs to use a weighted average"), it will sometimes quietly downgrade such a specialized requirement (for example, to a non-weighted average), without even creating an appropriate "WARNING:" comment in the generated code. Especially so when the relevant codebase already includes a similar, but not exactly appropriate API. My repetitive prompts to ALWAYS ask about ANY implementation ambiguities WHATSOEVER go unanswered.
From what I gather Claude Code seems to be better than other agents at always remembering to query the user about implementation ambiguities, so maybe I will give Claude Code a shot over Antigravity.
This all looks fine for someone who can't code, but for anyone with even a moderate amount of experience as a developer all this planning and checking and prompting and orchestrating is far more work than just writing the code yourself.
There's no winner for "least amount of code written regardless of productivity outcomes.", except for maybe Anthropic's bank account.
I really don't understand why there are so many comments like this.
Yesterday I had Claude write an audit logging feature to track all changes made to entities in my app. Yeah you get this for free with many frameworks, but my company's custom setup doesn't have it.
It took maybe 5-10 minutes of wall-time to come up with a good plan, and then ~20-30 min for Claude implement, test, etc.
That would've taken me at least a day, maybe two. I had 4-5 other tasks going on in other tabs while I waited the 20-30 min for Claude to generate the feature.
After Claude generated, I needed to manually test that it worked, and it did. I then needed to review the code before making a PR. In all, maybe 30-45 minutes of my actual time to add a small feature.
All I can really say is... are you sure you're using it right? Have you _really_ invested time into learning how to use AI tools?
Same here. I did bounce off these tools a year ago. They just didn't work for me 60% of the time. I learned a bit in that initial experience though and walked away with some tasks ChatGPT could replace in my workflow. Mainly replacing scripts and reviewing single files or functions.
Fast forward to today and I tried the tools again--specifically Claude Code--about a week ago. I'm blown away. I've reproduced some tools that took me weeks at full-time roles in a single day. This is while reviewing every line of code. The output is more or less what I'd be writing as a principal engineer.
> The output is more or less what I'd be writing as a principal engineer.
I certainly hope this is not true, because then you're not competent for that role. Claude Code writes an absolutely incredible amount of unecessary and superfluous comments, it's makes asinine mistakes like forgetting to update logic in multiple places. It'll gladly drop the entire database when changing column formats, just as an example.
Trust me I'm very impressed at the progress AI has made, and maybe we'll get to the point where everything is 100% correct all the time and better than any human could write. I'm skeptical we can get there with the LLM approach though.
The problem is LLMs are great at simple implementation, even large amounts of simple implementation, but I've never seen it develop something more than trivial correctly. The larger problem is it's very often subtly but hugely wrong. It makes bad architecture decisions, it breaks things in pursuit of fixing or implementing other things. You can tell it has no concept of the "right" way to implement something. It very obviously lacks the "senior developer insight".
Maybe you can resolve some of these with large amounts of planning or specs, but that's the point of my original comment - at what point is it easier/faster/better to just write the code yourself? You don't get a prize for writing the least amount of code when you're just writing specs instead.
This is exactly what the article is about. The tradeoff is that you have to throughly review the plans and iterate on them, which is tiring. But the LLM will write good code faster than you, if you tell it what good code is.
My experience has so far been similar to the root commenter - at the stage where you need to have a long cycle with planning it's just slower than doing the writing + theory building on my own.
It's an okay mental energy saver for simpler things, but for me the self review in an actual production code context is much more draining than writing is.
I guess we're seeing the split of people for whom reviewing is easy and writing is difficult and vice versa.
Exactly; the original commenter seems determined to write-off AI as "just not as good as me".
The original article is, to me, seemingly not that novel. Not because it's a trite example, but because I've begun to experience massive gains from following the same basic premise as the article. And I can't believe there's others who aren't using like this.
I iterate the plan until it's seemingly deterministic, then I strip the plan of implementation, and re-write it following a TDD approach. Then I read all specs, and generate all the code to red->green the tests.
If this commenter is too good for that, then it's that attitude that'll keep him stuck. I already feel like my projects backlog is achievable, this year.
Strongly agree about the deterministic part. Even more important than a good design, the plan must not show any doubt, whether it's in the form of open questions or weasel words. 95% of the time those vague words mean I didn't think something through, and it will do something hideous in order to make the plan work
Several months ago, just for fun, I asked Claude (the web site, not Claude Code) to build a web page with a little animated cannon that shoots at the mouse cursor with a ballistic trajectory. It built the page in seconds, but the aim was incorrect; it always shot too low. I told it the aim was off. It still got it wrong. I prompted it several times to try to correct it, but it never got it right. In fact, the web page started to break and Claude was introducing nasty bugs.
More recently, I tried the same experiment, again with Claude. I used the exact same prompt. This time, the aim was exactly correct. Instead of spending my time trying to correct it, I was able to ask it to add features. I've spent more time writing this comment on HN than I spent optimizing this toy. https://claude.ai/public/artifacts/d7f1c13c-2423-4f03-9fc4-8...
My point is that AI-assisted coding has improved dramatically in the past few months. I don't know whether it can reason deeply about things, but it can certainly imitate a human who reasons deeply. I've never seen any technology improve at this rate.
> but I've never seen it develop something more than trivial correctly.
What are you working on? I personally haven't seen LLMs struggle with any kind of problem in months. Legacy codebase with great complexity and performance-critical code. No issue whatsoever regardless of the size of the task.
Does it write maintainable code? Does it write extensible code? Does it write secure code? Does it write performant code?
My experience has been it failing most of these. The code might "work", but it's not good for anything more than trivial, well defined functions (that probably appeared in it's training data written by humans). LLMs have a fundamental lack of understanding of what they're doing, and it's obvious when you look at the finer points of the outcomes.
That said, I'm sure you could write detailed enough specs and provide enough examples to resolve these issues, but that's the point of my original comment - if you're just writing specs instead of code you're not gaining anything.
I find “maintainable code” the hardest bias to let go of. 15+ years of coding and design patterns are hard to let go.
But the aha moment for me was what’s maintainable by AI vs by me by hand are on different realms. So maintainable has to evolve from good human design patterns to good AI patterns.
Specs are worth it IMO. Not because if I can spec, I could’ve coded anyway. But because I gain all the insight and capabilities of AI, while minimizing the gotchas and edge failures.
> But the aha moment for me was what’s maintainable by AI vs by me by hand are on different realms
I don't find that LLMs are any more likely than humans to remember to update all of the places it wrote redundant functions. Generally far less likely, actually. So forgive me for treating this claim with a massive grain of salt.
> But the aha moment for me was what’s maintainable by AI vs by me by hand are on different realms. So maintainable has to evolve from good human design patterns to good AI patterns.
How do you square that with the idea that all the code still has to be reviewed by humans? Yourself, and your coworkers
I picture like semi conductors; the 5nm process is so absurdly complex that operators can't just peek into the system easily. I imagine I'm just so used to hand crafting code that I can't imagine not being able to peek in.
So maybe it's that we won't be reviewing by hand anymore? I.e. it's LLMs all the way down. Trying to embrace that style of development lately as unnatural as it feels. We're obv not 100% there yet but Claude Opus is a significant step in that direction and they keep getting better and better.
Then who is responsible when (not if) that code does horrible things? We have humans to blame right now. I just don’t see it happening personally because liability and responsibility are too important
And you don’t blame humans anyways lol. Everywhere I’ve worked has had “blameless” postmortems. You don’t remove human review unless you have reasonable alternatives like high test coverage and other automated reviews.
> Yesterday I had Claude write an audit logging feature to track all changes made to entities in my app. Yeah you get this for free with many frameworks, but my company's custom setup doesn't have it.
But did you truly think about such feature? Like guarantees that it should follow (like how do it should cope with entities migration like adding a new field) or what the cost of maintaining it further down the line. This looks suspiciously like drive-by PR made on open-source projects.
> That would've taken me at least a day, maybe two.
I think those two days would have been filled with research, comparing alternatives, questions like "can we extract this feature from framework X?", discussing ownership and sharing knowledge,.. Jumping on coding was done before LLMs, but it usually hurts the long term viability of the project.
Adding code to a project can be done quite fast (hackatons,...), ensuring quality is what slows things down in any any well functioning team.
I'll bite, because it does seem like something that should be quick in a well-architected codebase. What was the situation? Was there something in this codebase that was especially suited to AI-development? Large amounts of duplication perhaps?
I wanted to add audit logging for all endpoints we call, all places we call the DB, etc. across areas I haven't touched before. It would have taken me a while to track down all of the touchpoints.
Granted, I am not 100% certain that Claude didn't miss anything. I feel fairly confident that it is correct given that I had it research upfront, had multiple agents review, and it made the correct changes in the areas that I knew.
Also I'm realizing I didn't mention it included an API + UI for viewing events w/ pretty deltas
Well someone who says logging is easy never knows the difficulty of deciding "what" to log. And audit log is different beast altogether than normal logging
I'd find it deeply funny if the optimal vibe coding workflow continues to evolve to include more and more human oversight, and less and less agent autonomy, to the point where eventually someone makes a final breakthrough that they can save time by bypassing the LLM entirely and writing the code themselves. (Finally coming full circle.)
Researching and planning a project is a generally usefully thing. This is something I've been doing for years, and have always had great results compared to just jumping in and coding. It makes perfect sense that this transfers to LLM use.
Well it's less mental load. It's like Tesla's FSD. Am I a better driver than the FSD? For sure. But is it nice to just sit back and let it drive for a bit even if it's suboptimal and gets me there 10% slower, and maybe slightly pisses off the guy behind me? Yes, nice enough to shell out $99/mo. Code implementation takes a toll on you in the same way that driving does.
I think the method in TFA is overall less stressful for the dev. And you can always fix it up manually in the end; AI coding vs manual coding is not either-or.
Since Opus 4.5, things have changed quite a lot. I find LLMs very useful for discussing new features or ideas, and Sonnet is great for executing your plan while you grab a coffee.
Most of these AI coding articles seem to be about greenfield development.
That said, if you're on a serious team writing professional software there is still tons of value in always telling AI to plan first, unless it's a small quick task. This post just takes it a few steps further and formalizes it.
I find Cursor works much more reliably using plan mode, reviewing/revising output in markdown, then pressing build. Which isn't a ton of overhead but often leads to lots of context switching as it definitely adds more time.
I partly agree with you. But once you have a codebase large enough, the changes become longer to even type in, once figured out.
I find the best way to use agents (and I don't use claude) is to hash it out like I'm about to write these changes and I make my own mental notes, and get the agent to execute on it.
Agents don't get tired, they don't start fat fingering stuff at 4pm, the quality doesn't suffer. And they can be parallelised.
Finally, this allows me to stay at a higher level and not get bogged down of "right oh did we do this simple thing again?" which wipes some of the context in my mind and gets tiring through the day.
Always, 100% review every line of code written by an agent though. I do not condone committing code you don't 'own'.
I'll never agree with a job that forces developers to use 'AI', I sometimes like to write everything by hand. But having this tool available is also very powerful.
I want to be clear, I'm not against any use of AI. It's hugely useful to save a couple of minutes of "write this specific function to do this specific thing that I could write and know exactly what it would look like". That's a great use, and I use it all the time! It's better autocomplete. Anything beyond that is pushing it - at the moment! We'll see, but spending all day writing specs and double-checking AI output is not more productive than just writing correct code yourself the first time, even if you're AI-autocompleting some of it.
For the last few days I've been working on a personal project that's been on ice for at least 6 years. Back when I first thought of the project and started implementing it, it took maybe a couple weeks to eke out some minimally working code.
This new version that I'm doing (from scratch with ChatGPT web) has a far more ambitious scope and is already at the "usable" point. Now I'm primarily solidifying things and increasing test coverage. And I've tested the key parts with IRL scenarios to validate that it's not just passing tests; the thing actually fulfills its intended function so far. Given the increased scope, I'm guessing it'd take me a few months to get to this point on my own, instead of under a week, and the quality wouldn't be where it is. Not saying I haven't had to wrangle with ChatGPT on a few bugs, but after a decent initial planning phase, my prompts now are primarily "Do it"s and "Continue"s. Would've likely already finished it if I wasn't copying things back and forth between browser and editor, and being forced to pause when I hit the message limit.
I think it comes down to "it depends". I work in a NIS2 regulated field and we're quite callenged by the fact that it means we can't give AI's any sort of real access because of the security risk. To be complaint we'd have to have the AI agent ask permission for every single thing it does, before it does it, and foureye review it. Which is obviously never going to happen. We can discuss how bad the NIS2 foureye requirement works in the real world another time, but considering how easy it is to break AI security, it might not be something we can actually ever use. This makes sense on some of the stuff we work on, since it could bring an entire powerplant down. On the flip-side AI risks would be of little concern on a lot of our internal tools, which are basically non-regulated and unimportant enough that they can be down for a while without costing the business anything beyond annoyances.
This is where our challenges are. We've build our own chatbot where you can "build" your own agent within the librechat framework and add a "skill" to it. I say "skill" because it's older than claude skills but does exactly the same. I don't completely buy the authors:
> “deeply”, “in great details”, “intricacies”, “go through everything”
bit, but you can obviously save a lot of time by writing a piece of english which tells it what sort of environment you work in. It'll know that when I write Python I use UV, Ruff and Pyrefly and so on as an example. I personally also have a "skill" setting that tells the AI not to compliment me because I find that ridicilously annoying, and that certainly works. So who knows? Anyway, employees are going to want more. I've been doing some PoC's running open source models in isolation on a raspberry pi (we had spares because we use them in IoT projects) but it's hard to setup an isolation policy which can't be circumvented.
We'll have to figure it out though. For powerplant critical projects we don't want to use AI. But for the web tool that allows a couple of employees to upload three excel files from an external accountant and then generate some sort of report on them? Who cares who writes it or even what sort of quality it's written with? The lifecycle of that tool will probably be something that never changes until the external account does and then the tool dies. Not that it would have necessarily been written in worse quality without AI... I mean... Have you seen some of the stuff we've written in the past 40 years?
There is a miscommunication happening, this entire time we all had surprisingly different ideas about what quality of work is acceptable which seems to account for differences of opinion on this stuff.
> planning and checking and prompting and orchestrating is far more work than just writing the code yourself.
This! Once I'm familiar with the codebase (which I strive to do very quickly), for most tickets, I usually have a plan by the time I've read the description. I can have a couple of implementation questions, but I knew where the info is located in the codebase. For things, I only have a vague idea, the whiteboard is where I go.
The nice thing with such a mental plan, you can start with a rougher version (like a drawing sketch). Like if I'm starting a new UI screen, I can put a placeholder text like "Hello, world", then work on navigation. Once that done, I can start to pull data, then I add mapping functions to have a view model,...
Each step is a verifiable milestone. Describing them is more mentally taxing than just writing the code (which is a flow state for me). Why? Because English is not fit to describe how computer works (try describe a finite state machine like navigation flow in natural languages). My mental mental model is already aligned to code, writing the solution in natural language is asking me to be ambiguous and unclear on purpose.
Shameless plug: https://beadhub.ai allows you to do exactly that, but with several agents in parallel. One of them is in the role of planner, which takes care of the source-of-truth document and the long term view. They all stay in sync with real-time chat and mail.
The biggest roadblock to using agents to maximum effectiveness like this is the chat interface. It's convenience as detriment and convenience as distraction. I've found myself repeatedly giving into that convenience only to realize that I have wasted an hour and need to start over because the agent is just obliviously circling the solution that I thought was fully obvious from the context I gave it. Clearly these tools are exceptional at transforming inputs into outputs and, counterintuitively, not as exceptional when the inputs are constantly interleaved with the outputs like they are in chat mode.
What I've read is that even with all the meticulous planning, the author still needed to intervene. Not at the end but at the middle, unless it will continue building out something wrong and its even harder to fix once it's done. It'll cost even more tokens. It's a net negative.
You might say a junior might do the same thing, but I'm not worried about it, at least the junior learned something while doing that. They could do it better next time. They know the code and change it from the middle where it broke. It's a net positive.
Unfortunately, you could argue that the model provider has also learned something, i.e. the interaction can be used as additional training data to train subsequent models.
> One trick I use constantly: for well-contained features where I’ve seen a good implementation in an open source repo, I’ll share that code as a reference alongside the plan request. If I want to add sortable IDs, I paste the ID generation code from a project that does it well and say “this is how they do sortable IDs, write a plan.md explaining how we can adopt a similar approach.” Claude works dramatically better when it has a concrete reference implementation to work from rather than designing from scratch.
Licensing apparently means nothing.
Ripped off in the training data, ripped off in the prompt.
This is quite close to what I've arrived at, but with two modifications
1) anything larger I work on in layers of docs. Architecture and requirements -> design -> implementation plan -> code. Partly it helps me think and nail the larger things first, and partly helps claude. Iterate on each level until I'm satisfied.
2) when doing reviews of each doc I sometimes restart the session and clear context, it often finds new issues and things to clear up before starting the next phase.
Has anyone found a efficient way to avoid repeating the initial codebase assessment when working with large projects?
There are several projects on GitHub that attempt to tackle context and memory limitations, but I haven’t found one that consistently works well in practice.
My current workaround is to maintain a set of Markdown files, each covering a specific subsystem or area of the application. Depending on the task, I provide only the relevant documents to Claude Code to limit the context scope. It works reasonably well, but it still feels like a manual and fragile solution.
I’m interested in more robust strategies for persistent project context or structured codebase understanding.
Whenever I build a new feature with it I end up with several plan files leftover. I ask CC to combine them all, update with what we actually ended up building and name it something sensible, then whenever I want to work on that area again it's a useful reference (including the architecture, decisions and tradeoffs, relevant files etc).
For my longer spec files, I grep the subheaders/headers (with line numbers) and show this compact representation to the LLM's context window. I also have a file that describes what each spec files is and where it's located, and I force the LLM to read that and pull the subsections it needs. I also have one entrypoint requirements file (20k tokens) that I force it to read in full before it does anything else, every line I wrote myself. But none of this is a silver bullet.
That sounds like the recommended approach. However, there's one more thing I often do: whenever Claude Code and I complete a task that didn't go well at first, I ask CC what it learned, and then I tell it to write down what it learned for the future. It's hard to believe how much better CC has become since I started doing that. I ask it to write dozens of unit tests and it just does. Nearly perfectly. It's insane.
Skills almost seem like a solution, but they still need an out-of-band process to keep them updated as the codebase evolves. For now, a structured workflow that includes aggressive updates at the end of the loop is what I use.
I go a bit further than this and have had great success with 3 doc types and 2 skills:
- Specs: these are generally static, but updatable as the project evolves. And they're broken out to an index file that gives a project overview, a high-level arch file, and files for all the main modules. Roughly ~1k lines of spec for 10k lines of code, and try to limit any particular spec file to 300 lines. I'm intimately familiar with every single line in these.
- Plans: these are the output of a planning session with an LLM. They point to the associated specs. These tend to be 100-300 lines and 3 to 5 phases.
- Working memory files: I use both a status.md (3-5 items per phase roughly 30 lines overall), which points to a latest plan, and a project_status (100-200 lines), which tracks the current state of the project and is instructed to compact past efforts to keep it lean)
- A planner skill I use w/ Gemini Pro to generate new plans. It essentially explains the specs/plans dichotomy, the role of the status files, and to review everything in the pertinent areas of code and give me a handful of high-level next set of features to address based on shortfalls in the specs or things noted in the project_status file. Based on what it presents, I select a feature or improvement to generate. Then it proceeds to generate a plan, updates a clean status.md that points to the plan, and adjusts project_status based on the state of the prior completed plan.
- An implementer skill in Codex that goes to town on a plan file. It's fairly simple, it just looks at status.md, which points to the plan, and of course the plan points to the relevant specs so it loads up context pretty efficiently.
I've tried the two main spec generation libraries, which were way overblown, and then I gave superpowers a shot... which was fine, but still too much. The above is all homegrown, and I've had much better success because it keeps the context lean and focused.
And I'm only on the $20 plans for Codex/Gemini vs. spending $100/month on CC for half year prior and move quicker w/ no stall outs due to token consumption, which was regularly happening w/ CC by the 5th day. Codex rarely dips below 70% available context when it puts up a PR after an execution run. Roughly 4/5 PRs are without issue, which is flipped against what I experienced with CC and only using planning mode.
This is pretty much my approach. I started with some spec files for a project I'm working on right now, based on some academic papers I've written. I ended up going back and forth with Claude, building plans, pushing info back into the specs, expanding that out and I ended up with multiple spec/architecture/module documents. I got to the point where I ended up building my own system (using claude) to capture and generate artifacts, in more of a systems engineering style (e.g. following IEEE standards for conops, requirement documents, software definitions, test plans...). I don't use that for session-level planning; Claude's tools work fine for that. (I like superpowers, so far. It hasn't seemed too much)
I have found it to work very well with Claude by giving it context and guardrails. Basically I just tell it "follow the guidance docs" and it does. Couple that with intense testing and self-feedback mechanisms and you can easily keep Claude on track.
I have had the same experience with Codex and Claude as you in terms of token usage. But I haven't been happy with my Codex usage; Claude just feels like it's doing more of what I want in the way I want.
Looks good. Question - is it always better to use a monorepo in this new AI world? Vs breaking your app into separate repos? At my company we have like 6 repos all separate nextjs apps for the same user base. Trying to consolidate to one as it should make life easier overall.
It really depends but there’s nothing stopping you from just creating a separate folder with the cloned repositories (or worktrees) that you need and having a root CLAUDE.md file that explains the directory structure and referencing the individual repo CLAUDE.md files.
The multi-pass approach works outside of code too. I run a fairly complex automation pipeline (prompt -> script -> images -> audio -> video assembly) and the single biggest quality improvement was splitting generation into discrete planning and execution phases. One-shotting a 10-step pipeline means errors compound. Having the LLM first produce a structured plan, then executing each step against that plan with validation gates between them, cut my failure rate from maybe 40% to under 10%. The planning doc also becomes a reusable artifact you can iterate on without re-running everything.
> After Claude writes the plan, I open it in my editor and add inline notes directly into the document. These notes correct assumptions, reject approaches, add constraints, or provide domain knowledge that Claude doesn’t have.
This is the part that seems most novel compared to what I've heard suggested before. And I have to admit I'm a bit skeptical. Would it not be better to modify what Claude has written directly, to make it correct, rather than adding the corrections as separate notes (and expecting future Claude to parse out which parts were past Claude and which parts were the operator, and handle the feedback graciously)?
At least, it seems like the intent is to do all of this in the same session, such that Claude has the context of the entire back-and-forth updating the plan. But that seems a bit unpleasant; I would think the file is there specifically to preserve context between sessions.
One reason why I don't do this: even I won't be immune to mistakes. When I fix it with new values or paths, for example, and the one I provided is wrong, it can worsen the future work.
Personally, I like to order claude one more time to update the plan file after I have given annotation, and review it again after. This will ensure (from my understanding) that claude won't treat my annotation as different instructions, thus risking the work being conflicted.
I try these staging-document patterns, but suspect they have 2 fundamental flaws that stem mostly from our own biases.
First, Claude evolves. The original post work pattern evolved over 9 months, before claude's recent step changes. It's likely claude's present plan mode is better than this workaround, but if you stick to the workaround, you'd never know.
Second, the staging docs that represent some context - whether a library skills or current session design and implementation plans - are not the model Claude works with. At best they are shaping it, but I've found it does ignore and forget even what's written (even when I shout with emphasis), and the overall session influences the code. (Most often this happens when a peripheral adjustment ends up populating half the context.)
Indeed the biggest benefit from the OP might be to squeeze within 1 session, omitting peripheral features and investigations at the plan stage. So the mechanism of action might be the combination of getting our own plan clear and avoiding confusing excursions. (A test for that would be to redo the session with the final plan and implementation, to see if the iteration process itself is shaping the model.)
Our bias is to believe that we're getting better at managing this thing, and that we can control and direct it. It's uncomfortable to realize you can only really influence it - much like giving direction to a junior, but they can still go off track. And even if you found a pattern that works, it might work for reasons you're not understanding -- and thus fail you eventually. So, yes, try some patterns, but always hang on to the newbie senses of wonder and terror that make you curious, alert, and experimental.
This is the flow I've found myself working towards. Essentially maintaining more and more layered documentation for the LLM produces better and more consistent results. What is great here is the emphasis on the use of such documents in the planning phase. I'm feeling much more motivated to write solid documentation recently, because I know someone (the LLM) is actually going to read it! I've noticed my efforts and skill acquisition have moved sharply from app developer towards DevOps and architecture / management, but I think I'll always be grateful for the application engineering experience that I think the next wave of devs might miss out on.
I've also noted such a huge gulf between some developers describing 'prompting things into existence' and the approach described in this article. Both types seem to report success, though my experience is that the latter seems more realistic, and much more likely to produce robust code that's likely to be maintainable for long term or project critical goals.
The separation of planning and execution resonates strongly. I've been using a similar pattern when building with AI APIs — write the spec/plan in natural language first, then let the model execute against it.
One addition that's worked well for me: keeping a persistent context file that the model reads at the start of each session. Instead of re-explaining the project every time, you maintain a living document of decisions, constraints, and current state. Turns each session into a continuation rather than a cold start.
The biggest productivity gain isn't in the code generation itself — it's in reducing the re-orientation overhead between sessions.
Certainly the “unsupervised agent” workflows are getting a lot of attention right now, but they require a specific set of circumstances to be effective:
- clear validation loop (eg. Compile the kernel, here is gcc that does so correctly)
- ai enabled tooling (mcp / cli tool that will lint, test and provide feedback immediately)
- oversight to prevent sgents going off the rails (open area of research)
- an unlimited token budget
That means that most people can't use unsupervised agents.
Not that they dont work; Most people have simply not got an environment and task that is appropriate.
By comparison, anyone with cursor or claude can immediately start using this approach, or their own variant on it.
It does not require fancy tooling.
It does not require an arcane agent framework.
It works generally well across models.
This is one of those few genunie pieces of good practical advice for people getting into AI coding.
Simple. Obviously works once you start using it. No external dependencies. BYO tools to help with it, no “buy my AI startup xxx to help”. No “star my github so I can a job at $AI corp too”.
Honesty this is just language models in general at the moment, and not just coding.
It’s the same reason adding a thinking step works.
You want to write a paper, you have it form a thesis and structure first. (In this one you might be better off asking for 20 and seeing if any of them are any good.) You want to research something, first you add gathering and filtering steps before synthesis.
Adding smarter words or telling it to be deeper does work by slightly repositioning where your query ends up in space.
Asking for the final product first right off the bat leads to repetitive verbose word salad. It just starts to loop back in on itself. Which is why temperature was a thing in the first place, and leads me to believe they’ve turned the temp down a bit to try and be more accurate. Add some randomness and variability to your prompts to compensate.
Absolutely. And you can also always let the agent look back at the plan to check if it is still on track and aligned.
One step I added, that works great for me, is letting it write (api-level) tests after planning and before implementation. Then I’ll do a deep review and annotation of these tests and tweak them until everything is just right.
Can you help me understand the difference between "short prompt for what I want (next)" vs medium to high complexity tasks?
What i mean is, in practice, how does one even get to a a high complexity task? What does that look like? Because isn't it more common that one sees only so far ahead?
This is similar to what I do. I instruct an Architect mode with a set of rules related to phased implementation and detailed code artifacts output to a report.md file. After a couple of rounds of review and usually some responses that either tie together behaviors across context, critique poor choices or correct assumptions, there is a piece of work defined for a coder LLM to perform. With the new Opus 4.6 I then select specialist agents to review the report.md, prompted with detailed insight into particular areas of the software. The feedback from these specialist agent reviews is often very good and sometimes catches things I had missed. Once all of this is done, I let the agent make the changes and move onto doing something else. I typically rename and commit the report.md files which can be useful as an alternative to git diff / commit messages etc.
Good article, but I would rephrase the core principle slightly:
Never let Claude write code until you’ve reviewed, *fully understood* and approved a written plan.
In my experience, the beginning of chaos is the point at which you trust that Claude has understood everything correctly and claims to present the very best solution. At that point, you leave the driver's seat.
If you’ve ever desired the ability for annotating the plan more visually, try fitting Plannotator in this workflow. There is a slash command for use when you use custom workflows outside of normal plan mode.
The crowd around this pot shows how superficial is knowledge about claude code. It gets releases each day and most of this is already built in the vanilla version. Not to mention subagent working in work trees, memory.md, plan on which you can comment directly from the interface, subagents launched in research phase, but also some basic mcp's like LSP/IDE integration, and context7 to not to be stuck in the knowledge cutoff/past.
When you go to YouTube and search for stuff like "7 levels of claude code" this post would be maybe 3-4.
Oh, one more thing - quality is not consistent, so be ready for 2-3 rounds of "are you happy with the code you wrote" and defining audit skills crafted for your application domain - like for example RODO/Compliance audit etc.
I'm using the in-built features as well, but I like the flow that I have with superpowers. You've made a lot of assumptions with your comment that are just not true (at least for me).
I find that brainstorming + (executing plans OR subagent driven development) is way more reliable than the built-in tooling.
I do something very similar, also with Claude and Codex, because the workflow is controlled by me, not by the tool. But instead of plan.md I use a ticket system basically like ticket_<number>_<slug>.md where I let the agent create the ticket from a chat, correct and annotate it afterwards and send it back, sometimes to a new agent instance. This workflow helps me keeping track of what has been done over time in the projects I work on. Also this approach does not need any „real“ ticket system tooling/mcp/skill/whatever since it works purely on text files.
+1 to creating tickets by simply asking the agent to. It's worked great and larger tasks can be broken down into smaller subtasks that could reasonably be completed in a single context window, so you rarely every have to deal with compaction. Especially in the last few months since Claude's gotten good at dispatching agents to handle tasks if you ask it to, I can plan large changes that span multilpe tickets and tell claude to dispatch agents as needed to handle them (which it will do in parallel if they mostly touch different files), keeping the main chat relatively clean for orchestration and validation work.
I've been working off and on on a vibe coded FP language and transpiler - mostly just to get more experience with Claude Code and see how it handles complex real world projects. I've settled on a very similar flow, though I use three documents: plan, context, task list. Multiple rounds of iteration when planning a feature. After completion, have a clean session do an audit to confirm that everything was implemented per the design. Then I have both Claude and CodeRabbit do code review passes before I finally do manual review. VERY heavy emphasis on tests, the project currently has 2x more test code than application code. So far it works surprisingly well. Example planning docs below -
The annotation cycle is the key insight for me. Treating the plan as a living doc you iterate on before touching any code makes a huge difference in output quality.
Experimentally, i've been using mfbt.ai [https://mfbt.ai] for roughly the same thing in a team context. it lets you collaboratively nail down the spec with AI before handing off to a coding agent via MCP.
Avoids the "everyone has a slightly different plan.md on their machine" problem. Still early days but it's been a nice fit for this kind of workflow.
I agree, and this is why I tend to use gptel in emacs for planning - the document is the conversation context, and can be edited and annotated as you like.
I've been teaching AI coding tool workshops for the past year and this planning-first approach is by far the most reliable pattern I've seen across skill levels.
The key insight that most people miss: this isn't a new workflow invented for AI - it's how good senior engineers already work. You read the code deeply, write a design doc, get buy-in, then implement. The AI just makes the implementation phase dramatically faster.
What I've found interesting is that the people who struggle most with AI coding tools are often junior devs who never developed the habit of planning before coding. They jump straight to "build me X" and get frustrated when the output is a mess. Meanwhile, engineers with 10+ years of experience who are used to writing design docs and reviewing code pick it up almost instantly - because the hard part was always the planning, not the typing.
One addition I'd make to this workflow: version your research.md and plan.md files in git alongside your code. They become incredibly valuable documentation for future maintainers (including future-you) trying to understand why certain architectural decisions were made.
I have to give this a try. My current model for backend is the same as how author does frontend iteration. My friend does the research-plan-edit-implement loop, and there is no real difference between the quality of what I do and what he does. But I do like this just for how it serves as documentation of the thought process across AI/human, and can be added to version control. Instead of humans reviewing PRs, perhaps humans can review the research/plan document.
On the PR review front, I give Claude the ticket number and the branch (or PR) and ask it to review for correctness, bugs and design consistency. The prompt is always roughly the same for every PR. It does a very good job there too.
The author is quite far on their journey but would benefit from writing simple scripts to enforce invariants in their codebase. Invariant broken? Script exits with a non-zero exit code and some output that tells the agent how to address the problem. Scripts are deterministic, run in milliseconds, and use zero tokens. Put them in husky or pre-commit, install the git hooks, and your agent won’t be able to commit without all your scripts succeeding.
And “Don’t change this function signature” should be enforced not by anticipating that your coding agent “might change this function signature so we better warn it not to” but rather via an end to end test that fails if the function signature is changed (because the other code that needs it not to change now has an error). That takes the author out of the loop and they can not watch for the change in order to issue said correction, and instead sip coffee while the agent observes that it caused a test failure then corrects it without intervention, probably by rolling back the function signature change and changing something else.
Haha this is surprisingly and exactly how I use claude as well. Quite fascinating that we independently discovered the same workflow.
I maintain two directories: "docs/proposals" (for the research md files) and "docs/plans" (for the planning md files). For complex research files, I typically break them down into multiple planning md files so claude can implement one at a time.
A small difference in my workflow is that I use subagents during implementation to avoid context from filling up quickly.
Same, I formalized a similar workflow for my team (oriented around feature requirement docs), I am thinking about fully productizing it and am looking to for feedback - https://acai.sh
Even if the product doesn’t resonate I think I’ve stumbled on some ideas you might find useful^
I do think spec-driven development is where this all goes. Still making up my mind though.
Spec-driven looks very much like what the author describes. He may have some tweaks of his own but they could just as well be coded into the artifacts that something like OpenSpec produces.
This is basically long-lived specs that are used as tests to check that the product still adheres to the original idea that you wanted to implement, right?
This inspired me to finally write good old playwright tests for my website :).
I don't deny that AI has use cases, but boy - the workflow described is boring:
"Most developers type a prompt, sometimes use plan mode, fix the errors, repeat. "
Does anyone think this is as epic as, say, watch the Unix archives https://www.youtube.com/watch?v=tc4ROCJYbm0 where Brian demos how pipes work; or Dennis working on C and UNIX? Or even before those, the older machines?
I am not at all saying that AI tools are all useless, but there is no real epicness. It is just autogenerated AI slop and blob. I don't really call this engineering (although I also do agree, that it is engineering still; I just don't like using the same word here).
> never let Claude write code until you’ve reviewed and approved a written plan.
So the junior-dev analogy is quite apt here.
I tried to read the rest of the article, but I just got angrier. I never had that feeling watching oldschool legends, though perhaps some of their work may be boring, but this AI-generated code ... that's just some mythical random-guessing work. And none of that is "intelligent", even if it may appear to work, may work to some extent too. This is a simulation of intelligence. If it works very well, why would any software engineer still be required? Supervising would only be necessary if AI produces slop.
Interesting! I feel like I'm learning to code all over again! I've only been using Claude for a little more than a month and until now I've been figuring things out on my own. Building my methodology from scratch. This is much more advanced than what I'm doing. I've been going straight to implementation, but doing one very small and limited feature at a time, describing implementation details (data structures like this, use that API here, import this library etc) verifying it manually, and having Claude fix things I don't like. I had just started getting annoyed that it would make the same (or very similar) mistake over and over again and I would have to fix it every time. This seems like it'll solve that problem I had only just identified! Neat!
> Most developers type a prompt, sometimes use plan mode, fix the errors, repeat.
> ...
> never let Claude write code until you’ve reviewed and approved a written plan
I certainly always work towards an approved plan before I let it lost on changing the code. I just assumed most people did, honestly. Admittedly, sometimes there's "phases" to the implementation (because some parts can be figured out later and it's more important to get the key parts up and running first), but each phase gets a full, reviewed plan before I tell it to go.
In fact, I just finished writing a command and instruction to tell claude that, when it presents a plan for implementation, offer me another option; to write out the current (important parts of the) context and the full plan to individual (ticket specific) md files. That way, if something goes wrong with the implementation I can tell it to read those files and "start from where they left off" in the planning.
I recently discovered GitHub speckit which separates planning/execution in stages: specify, plan, tasks, implement. Finding it aligns with the OP with the level of “focus” and “attention” this gets out of Claude Code.
Speckit is worth trying as it automates what is being described here, and with Opus 4.6 it's been a kind of BC/AD moment for me.
Try OpenSpec and it'll do all this for you. SpecKit works too. I don't think there's a need to reinvent the wheel on this one, as this is spec-driven development.
Since the rise of AI systems I really wonder how people wrote code before. This is exactly how I planned out implementation and executed the plan. Might have been some paper notes, a ticket or a white board, buuuuut ... I don't know.
I tried Opus 4.6 recently and it’s really good. I had ditched Claude a long time ago for Grok + Gemini + OpenCode with Chinese models. I used Grok/Gemini for planning and core files, and OpenCode for setup, running, deploying, and editing.
However, Opus made me rethink my entire workflow. Now, I do it like this:
* PRD (Product Requirements Document)
* main.py + requirements.txt + readme.md (I ask for minimal, functional, modular code that fits the main.py)
* Ask for a step-by-step ordered plan
* Ask to focus on one step at a time
The super powerful thing is that I don’t get stuck on missing accounts, keys, etc. Everything is ordered and runs smoothly. I go rapidly from idea to working product, and it’s incredibly easy to iterate if I figure out new features are required while testing. I also have GLM via OpenCode, but I mainly use it for "dumb" tasks.
Interestingly, for reasoning capabilities regarding standard logic inside the code, I found Gemini 3 Flash to be very good and relatively cheap. I don't use Claude Code for the actual coding because forcing everything via chat into a main.py encourages minimal code that's easy to skim—it gives me a clearer representation of the feature space
Interesting approach. The separation of planning and execution is crucial, but I think there's a missing layer most people overlook: permission boundaries between the two phases.
Right now when Claude Code (or any agent) executes a plan, it typically has the same broad permissions for every step. But ideally, each execution step should only have access to the specific tools and files it needs — least privilege, applied to AI workflows.
I've been experimenting with declarative permission manifests for agent tasks. Instead of giving the agent blanket access, you define upfront what each skill can read, write, and execute. Makes the planning phase more constrained but the execution phase much safer.
Anyone else thinking about this from a security-first angle?
I’m a big fan of having the model create a GitHub issue directly (using the GH CLI) with the exact plan it generates, instead of creating a markdown file that will eventually get deleted. It gives me a permanent record and makes it easy to reference and close the issue once the PR is ready.
Google Anti-Gravity has this process built in. This is essentially a cycle a developer would follow: plan/analyse - document/discuss - break down tasks/implement. We’ve been using requirements and design documents as best practice since leaving our teenage bedroom lab for the professional world. I suppose this could be seen as our coding agents coming of age.
I came to the exact same pattern, with one extra heuristic at the end: spin up a new claude instance after the implementation is complete and ask it to find discrepancies between the plan and the implementation.
> I am not seeing the performance degradation everyone talks about after 50% context window.
I pretty much agree with that. I use long sessions and stopped trying to optimize the context size, the compaction happens but the plan keeps the details and it works for me.
I agree with most of this, though I'm not sure it's radically different. I think most people who've been using CC in earnest for a while probably have a similar workflow? Prior to Claude 4 it was pretty much mandatory to define requirements and track implementation manually to manage context. It's still good, but since 4.5 release, it feels less important. CC basically works like this by default now, so unless you value the spec docs (still a good reference for Claude, but need to be maintained), you don't have to think too hard about it anymore.
The important thing is to have a conversation with Claude during the planning phase and don't just say "add this feature" and take what you get. Have a back and forth, ask questions about common patterns, best practices, performance implications, security requirements, project alignment, etc. This is a learning opportunity for you and Claude. When you think you're done, request a final review to analyze for gaps or areas of improvement. Claude will always find something, but starts to get into the weeds after a couple passes.
If you're greenfield and you have preferences about structure and style, you need to be explicit about that. Once the scaffolding is there, modern Claude will typically follow whatever examples it finds in the existing code base.
I'm not sure I agree with the "implement it all without stopping" approach and let auto-compact do its thing. I still see Claude get lazy when nearing compaction, though has gotten drastically better over the last year. Even so, I still think it's better to work in a tight loop on each stage of the implementation and preemptively compacting or restarting for the highest quality.
Not sure that the language is that important anymore either. Claude will explore existing codebase on its own at unknown resolution, but if you say "read the file" it works pretty well these days.
My suggestions to enhance this workflow:
- If you use a numbered phase/stage/task approach with checkboxes, it makes it easy to stop/resume as-needed, and discuss particular sections. Each phase should be working/testable software.
- Define a clear numbered list workflow in CLAUDE.md that loops on each task (run checks, fix issues, provide summary, etc).
- Use hooks to ensure the loop is followed.
- Update spec docs at the end of the cycle if you're keeping them. It's not uncommon for there to be some divergence during implementation and testing.
I don't really get what is different about this from how almost everyone else uses Claude Code? This is an incredibly common, if not the most common way of using it (and many other tools).
Another approach is to spec functionality using comments and interfaces, then tell the LLM to first implement tests and finally make the tests pass. This way you also get regression safety and can inspect that it works as it should via the tests.
this sounds... really slow. for large changes for sure i'm investing time into planning. but such a rigid system can't possible be as good as a flexible approach with variable amounts of planning based on complexity
Insights are nice for new users but I’m not seeing anything too different from how anyone experienced with Claude Code would use plan mode. You can reject plans with feedback directly in the CLI.
It is really fun to watch how a baby makes its first steps and also how experienced professionals rediscover what standards were telling us for 80+ years.
Funny how I came up with something loosely similar. Asking Codex to write a detailed plan in a markdown document, reviewing it, and asking it to implement it step by step. It works exquisitely well when it can build and test itself.
What works extremely well for me is this: Let Claude Code create the plan, then turn over the plan to Codex for review, and give the response back to Claude Code. Codex is exceptionally good at doing high level reviews and keeping an eye on the details. It will find very suble errors and omissins. And CC is very good at quickly converting the plan into code.
This back and forth between the two agents with me steering the conversation elevates Claude Code into next level.
Hub and spoke documentation in planning has been absolutely essential for the way my planning was before, and it's pretty cool seeing it work so well for planning mode to build scaffolds and routing.
this is literally reinventing claude's planning mode, but with more steps. I think Boris doesn't realize that planning mode is actually stored in a file.
I have tried using this and other workflows for a long time and had never been able to get them to work (see chat history for details).
This has changed in the last week, for 3 reasons:
1. Claude opus. It’s the first model where I haven’t had to spend more time correcting things than it would’ve taken me to just do it myself. The problem is that opus chews through tokens, which led to..
2. I upgraded my Claude plan. Previously on the regular plan I’d get about 20 mins of time before running out of tokens for the session and then needing to wait a few hours to use again. It was fine for little scripts or toy apps but not feasible for the regular dev work I do. So I upgraded to 5x. This now got me 1-2 hours per session before tokens expired. Which was better but still a frustration. Wincing at the price, I upgraded again to the 20x plan and this was the next game changer. I had plenty of spare tokens per session and at that price it felt like they were being wasted - so I ramped up my usage. Following a similar process as OP but with a plans directory with subdirectories for backlog, active and complete plans, and skills with strict rules for planning, implementing and completing plans, I now have 5-6 projects on the go. While I’m planning a feature on one the others are implementing. The strict plans and controls keep them on track and I have follow up skills for auditing quality and performance. I still haven’t hit token limits for a session but I’ve almost hit my token limit for the week so I feel like I’m getting my money’s worth. In that sense spending more has forced me to figure out how to use more.
3. The final piece of the puzzle is using opencode over claude code. I’m not sure why but I just don’t gel with Claude code. Maybe it’s all the sautéing and flibertygibbering, maybe it’s all the permission asking, maybe it’s that it doesn’t show what it’s doing as much as opencode. Whatever it is it just doesn’t work well for me. Opencode on the other hand is great. It’s shows what it’s doing and how it’s thinking which makes it easy for me to spot when it’s going off track
and correct early.
Having a detailed plan, and correcting and iterating on the plan is essential. Making clause follow the plan is also essential - but there’s a line. Too fine grained and it’s not as creative at solving problems. Too loose/high level and it makes bad choices and goes in the wrong direction.
Is it actually making me more productive? I think it is but I’m only a week in. I’ve decided to give myself a month to see how it all works out.
I don’t intend to keep paying for the 20x plan unless I can see a path to using it to earn me at least as much back.
It isn’t slower. I use my personal ChatGPT subscriptions with Codex for almost everything at work and use my $800/month company Claude allowance only for the tricky stuff that Codex can’t figure out. It’s never application code. It’s usually some combination of app code + Docker + AWS issue with my underlying infrastructure - created with whatever IAC that I’m using for a client - Terraform/CloudFormation or the CDK.
I burned through $10 on Claude in less than an hour. I only have $36 a day at $800 a month (800/22 working days)
I use both. As I’m working, I tell each of them to update a common document with the conversation. I don’t just tell Claude the what. I tell it the why and have it document it.
I can switch back and forth and use the MD file as shared context.
Curious: what are some cases where it'd make sense to not pay for the 20x plan (which is $200/month), and provide a whopping $800/month pay-per-token allowance instead?
Who knows? It’s part of an enterprise plan. I work for a consulting company. There are a number of fallbacks, the first fallback if we are working on an internal project is just to use our internal AWS account and use Claude code with the Anthropic hosted on Bedrock.
The second fallback if it is for a customer project is to use their AWS account for development for them.
The rate my company charges for me - my level as an American based staff consultant (highest bill rate at the company) they are happy to let us use Claude Code using their AWS credentials. Besides, if we are using AWS Bedrock hosted Anthropic models, they know none of their secrets are going to Anthropic. They already have the required legal confidentiality/compliancd agreements with AWS.
It’s worrying to me that nobody really knows how LLMs work. We create prompts with or without certain words and hope it works. That’s my perspective anyway
It's actually no different from how real software is made. Requirements come from the business side, and through an odd game of telephone get down to developers.
The team that has developers closest to the customer usually makes the better product...or has the better product/market fit.
It's the same as dealing with a human. You convey a spec for a problem and the language you use matters. You can convey the problem in (from your perspective) a clear way and you will get mixed results nonetheless. You will have to continue to refine the solution with them.
Genuinely: no one really knows how humans work either.
Is it required to tell Claude to re-read the code folder again when you come back some day later or should we ask Claude to just pickup from research.md file thus saving some tokens?
Sorry but I didn't get the hype with this post, isnt it what most of the people doing? I want to see more posts on how you use the claude "smart" without feeding the whole codebase polluting the context window and also more best practices on cost efficient ways to use it, this workflow is clearly burning million tokens per session, for me is a No
I do something broadly similar. I ask for a design doc that contains an embedded todo list, broken down into phases. Looping on the design doc asking for suggestions seems to help. I'm up to about 40 design docs so far on my current project.
That's exactly what Cursor's "plan" mode does? It even creates md files, which seems to be the main "thing" the author discovered. Along with some cargo cult science?
How is this noteworthy other than to spark a discussion on hn? I mean I get it, but a little more substance would be nice.
This is great. My workflow is also heading in that direction, so this is a great roadmap. I've already learned that just naively telling Claude what to do and letting it work, is a recipe for disaster and wasted time.
I'm not this structured yet, but I often start with having it analyse and explain a piece of code, so I can correct it before we move on. I also often switch to an LLM that's separate from my IDE because it tends to get confused by sprawling context.
* I ask the LLM for it's understanding of a topic or an existing feature in code. It's not really planning, it's more like understanding the model first
* Then based on its understanding, I can decide how great or small to scope something for the LLM
* An LLM showing good understand can deal with a big task fairly well.
* An LLM showing bad understanding still needs to be prompted to get it right
* What helps a lot is reference implementations. Either I have existing code that serves as the reference or I ask for a reference and I review.
A few folks do it at my work do it OPs way, but my arguments for not doing it this way
* Nobody is measuring the amount of slop within the plan. We only judge the implementation at the end
* it's still non deterministic - folks will have different experiences using OPs methods. If claude updates its model, it outdates OPs suggestions by either making it better or worse. We don't evaluate when things get better, we only focus on things not gone well.
* it's very token heavy - LLM providers insist that you use many tokens to get the task done. It's in their best interest to get you to do this. For me, LLMs should be powerful enough to understand context with minimal tokens because of the investment into model training.
Both ways gets the task done and it just comes down to my preference for now.
For me, I treat the LLM as model training + post processing + input tokens = output tokens. I don't think this is the best way to do non deterministic based software development. For me, we're still trying to shoehorn "old" deterministic programming into a non deterministic LLM.
Tip:
LLMs are very good at following conventions (this is actually what is happening when it writes code).
If you create a .md file with a list of entries of the following structure:
# <identifier>
<description block>
<blank space>
# <identifier>
...
where an <identifier> is a stable and concise sequence of tokens that identifies some "thing" and seed it with 5 entries describing abstract stuff, the LLM will latch on and reference this. I call this a PCL (Project Concept List). I just tell it:
> consume tmp/pcl-init.md pcl.md
The pcl-init.md describes what PCL is and pcl.md is the actual list.
I have pcl.md file for each independent component in the code (logging, http, auth, etc).
This works very very well.
The LLM seems to "know" what you're talking about.
You can ask questions and give instructions like "add a PCL entry about this".
It will ask if should add a PCL entry about xyz.
If the description block tends to be high information-to-token ratio, it will follow that convention (which is a very good convention BTW).
However, there is a caveat. LLMs resist ambiguity about authority. So the "PCL" or whatever you want to call it, needs to be the ONE authoritative place for everything. If you have the same stuff in 3 different files, it won't work nearly as well.
Bonus Tip:
I find long prompt input with example code fragments and thoughtful descriptions work best at getting an LLM to produce good output. But there will always be holes (resource leaks, vulnerabilities, concurrency flaws, etc). So then I update my original prompt input (keep it in a separate file PROMPT.txt as a scratch pad) to add context about those things maybe asking questions along the way to figure out how to fix the holes. Then I /rewind back to the prompt and re-enter the updated prompt. This feedback loop advances the conversation without expending tokens.
I don't know. I tried various methods. And this one kind of doesn't work quite a bit of times. The problem is plan naturally always skips some important details, or assumes some library function, but is taken as instruction in the next section. And claude can't handle ambiguity if the instruction is very detailed(e.g. if plan asks to use a certain library even if it is a bad fit claude won't know that decision is flexible). If the instruction is less detailed, I saw claude is willing to try multiple things and if it keeps failing doesn't fear in reverting almost everything.
In my experience, the best scenario is that instruction and plan should be human written, and be detailed.
I have a different approach where I have claude write coding prompts for stages then I give the prompt to another agent. I wonder if I should write it up as a blog post
add another agent review, I ask Claude to send plan for review to Codex and fix critical and high issues, with complexity gating (no overcomplicated logic), run in a loop, then send to Gemini reviewer, then maybe final pass with Claude, once all C+H pass the sequence is done
It looks verbose but it defines the requirements based on your input, and when you approve it then it defines a design, and (again) when you approve it then it defines an implementation plan (a series of tasks.)
This separation of planning and execution resonates deeply with how I approach task management in general, not just coding.
The key insight here - that planning and execution should be distinct phases - applies to productivity tools too. I've been using www.dozy.site which takes a similar philosophy: it has smart calendar scheduling that automatically fills your empty time slots with planned tasks. The planning happens first (you define your tasks and projects), then the execution is automated (tasks get scheduled into your calendar gaps).
The parallel is interesting: just like you don't want Claude writing code before the plan is solid, you don't want to manually schedule tasks before you've properly planned what needs to be done. The separation prevents wasted effort and context switching.
The annotation cycle you describe (plan -> review -> annotate -> refine) is exactly how I work with my task lists too. Define the work, review it, adjust priorities and dependencies, then let the system handle the scheduling.
Almost think we're at the point on HN where we need a special [flag bot] link for those that meet a certain threshold and it alerts @dang or something to investigate them in more detail. The amount of bots on here has been increasing at an alarming rate.
There has been this really weird flood of new accounts lately that are making these kinds of bot comments with no clear purpose to making them. Maybe it comes from people experimenting with OpenClaw?
I don't see how this is 'radically different' given that Claude Code literally has a planning mode.
This is my workflow as well, with the big caveat that 80% of 'work' doesn't require substantive planning, we're making relatively straight forward changes.
Edit: there is nothing fundamentally different about 'annotating offline' in an MD vs in the CLI and iterating until the plan is clear. It's a UI choice.
Spec Driven Coding with AI is very well established, so working from a plan, or spec (they can be somewhat different) is not novel.
last i checked, you can't annotate inline with planning mode. you have to type a lot to explain precisely what needs to change, and then it re-presents you with a plan (which may or may not have changed something else).
i like the idea of having an actual document because you could actually compare the before and after versions if you wanted to confirm things changed as intended when you gave feedback
Wow, I've been needing this! The one issue I’ve had with terminals is reviewing plans, and desiring the ability to provide feedback on specific plan sections in a more organized way.
I just use Jessie’s “superpowers” plugin. It does all of this but also steps you through the design and gives you bite sized chunks and you make architecture decisions along the way. Far better than making big changes to an already established plan.
The author seems to think they've hit upon something revolutionary...
They've actually hit upon something that several of us have evolved to naturally.
LLM's are like unreliable interns with boundless energy. They make silly mistakes, wander into annoying structural traps, and have to be unwound if left to their own devices. It's like the genie that almost pathologically misinterprets your wishes.
So, how do you solve that? Exactly how an experienced lead or software manager does: you have systems write it down before executing, explain things back to you, and ground all of their thinking in the code and documentation, avoiding making assumptions about code after superficial review.
When it was early ChatGPT, this meant function-level thinking and clearly described jobs. When it was Cline it meant cline rules files that forced writing architecture.md files and vibe-code.log histories, demanding grounding in research and code reading.
Maybe nine months ago, another engineer said two things to me, less than a day apart:
- "I don't understand why your clinerules file is so large. You have the LLM jumping through so many hoops and doing so much extra work. It's crazy."
- The next morning: "It's basically like a lottery. I can't get the LLM to generate what I want reliably. I just have to settle for whatever it comes up with and then try again."
These systems have to deal with minimal context, ambiguous guidance, and extreme isolation. Operate with a little empathy for the energetic interns, and they'll uncork levels of output worth fighting for. We're Software Managers now. For some of us, that's working out great.
Revolutionary or not it was very nice of the author to make time and effort to share their workflow.
For those starting out using Claude Code it gives a structured way to get things done bypassing the time/energy needed to “hit upon something that several of us have evolved to naturally”.
It's this line that I'm bristling at: "...the workflow I’ve settled into is radically different from what most people do with AI coding tools..."
Anyone who spends some time with these tools (and doesn't black out from smashing their head against their desk) is going to find substantial benefit in planning with clarity.
It was #6 in Boris's run-down: https://news.ycombinator.com/item?id=46470017
So, yes, I'm glad that people write things out and share. But I'd prefer that they not lead with "hey folks, I have news: we should *slice* our bread!"
But the author's workflow is actually very different from Boris'.
#6 is about using plan mode whereas the author says "The built-in plan mode sucks".
The author's post is much more than just "planning with clarity".
Its ai written though, the tells are in pretty much every paragraph.
I don’t think it’s that big a red flag anymore. Most people use ai to rewrite or clean up content, so I’d think we should actually evaluate content for what it is rather than stop at “nah it’s ai written.”
Even though I use LLMs for code, I just can't read LLM written text, I kind of hate the style, it reminds me too much of LinkedIn.
I don't judge content for being AI written, I judge it for the content itself (just like with code).
However I do find the standard out-of-the-box style very grating. Call it faux-chummy linkedin corporate workslop style.
Why don't people give the llm a steer on style? Either based on your personal style or at least on a writer whose style you admire. That should be easier.
Because they think this is good writing. You can’t correct what you don’t have taste for. Most software engineers think that reading books means reading NYT non-fiction bestsellers.
>Most people use ai to rewrite or clean up content
I think your sentence should have been "people who use ai do so to mostly rewrite or clean up content", but even then I'd question the statistical truth behind that claim.
Personally, seeing something written by AI means that the person who wrote it did so just for looks and not for substance. Claiming to be a great author requires both penmanship and communication skills, and delegating one or either of them to a large language model inherently makes you less than that.
However, when the point is just the contents of the paragraph(s) and nothing more then I don't care who or what wrote it. An example is the result of a research, because I'd certainly won't care about the prose or effort given to write the thesis but more on the results (is this about curing cancer now and forever? If yes, no one cares if it's written with AI).
With that being said, there's still that I get anywhere close to understanding the author behind the thoughts and opinions. I believe the way someone writes hints to the way they think and act. In that sense, using LLM's to rewrite something to make it sound more professional than what you would actually talk in appropriate contexts makes it hard for me to judge someone's character, professionalism, and mannerisms. Almost feels like they're trying to mask part of themselves. Perhaps they lack confidence in their ability to sound professional and convincing?
Well, real humans may read it though. Personally I much prefer real humans write real articles than all this AI generated spam-slop. On youtube this is especially annoying - they mix in real videos with fake ones. I see this when I watch animal videos - some animal behaviour is taken from older videos, then AI fake is added. My own policy is that I do not watch anything ever again from people who lie to the audience that way so I had to begin to censor away such lying channels. I'd apply the same rationale to blog authors (but I am not 100% certain it is actually AI generated; I just mention this as a safety guard).
If you want to write something with AI, send me your prompt. I'd rather read what you intend for it to produce rather than what it produces. If I start to believe you regularly send me AI written text, I will stop reading it. Even at work. You'll have to call me to explain what you intended to write.
And if my prompt is a 60 page wall of text that I would otherwise take the time to have the AI organize and summarize and sharpen with an index, executive summary, and logical sections, are you going to actually read all of that, or just whine "TL;DR"?
It's much more efficient and intentional for the writer to do the condensing and organizing once, and review and proofread it to make sure it's what they mean, than for the author to just lazily spam every human they want to read it with the raw prompt, so every recipient has to pay for their own AI to perform that task, producing random results not reviewed and approved by the author as their intended message.
Is that really how you want Hacker News discussions and your work email to be, walls of unorganized unfiltered text prompts nobody including yourself wants to take the time to read? Then step aside, hold my beer!
The main issue with evaluating content for what it is is how extremely asymmetric that process has become.
Slop looks reasonable on the surface, and requires orders of magnitude more effort to evaluate than to produce. It’s produced once, but the process has to be repeated for every single reader.
Disregarding content that smells like AI becomes an extremely tempting early filtering mechanism to separate signal from noise - the reader’s time is valuable.
I think as humans it's very hard to abstract content from its form. So when the form is always the same boring, generic AI slop, it's really not helping the content.
And maybe writing an article or a keynote slides is one of the few places we can still exerce some human creativity, especially when the core skills (programming) is almost completely in the hands of LLMs already
> I don’t think it’s that big a red flag anymore.
It is to me, because it indicates the author didn't care about the topic. The only thing they cared about is to write an "insightful" article about using llms. Hence this whole thing is basically linked-in resume improvement slop.
Not worth interacting with, imo
Also, it's not insightful whatsoever. It's basically a retelling of other articles around the time Claude code was released to the public (March-August 2025)
>the tells are in pretty much every paragraph.
It's not just misleading — it's lazy. And honestly? That doesn't vibe with me.
[/s obviously]
So is GP.
This is clearly a standard AI exposition:
LLM's are like unreliable interns with boundless energy. They make silly mistakes, wander into annoying structural traps, and have to be unwound if left to their own devices. It's like the genie that almost pathologically misinterprets your wishes.
Then ask your own ai to rewrite it so it doesn't trigger you into posting uninteresting thought stopping comments proclaiming why you didn't read the article, that don't contribute to the discussion.
Here's mine! https://github.com/pjlsergeant/moarcode
It feels like retracing the history of software project management. The post is quite waterfall-like. Writing a lot of docs and specs upfront then implementing. Another approach is to just YOLO (on a new branch) make it write up the lessons afterwards, then start a new more informed try and throw away the first. Or any other combo.
For me what works well is to ask it to write some code upfront to verify its assumptions against actual reality, not just be telling it to review the sources "in detail". It gains much more from real output from the code and clears up wrong assumptions. Do some smaller jobs, write up md files, then plan the big thing, then execute.
This is exactly what I do. I assume most people avoid this approach due to cost.
It makes an endless stream of assumptions. Some of them brilliant and even instructive to a degree, but most of them are unfounded and inappropriate in my experience.
Agreed. The process described is much more elaborate than what I do but quite similar. I start to discuss in great details what I want to do, sometimes asking the same question to different LLMs. Then a todo list, then manual review of the code, esp. each function signature, checking if the instructions have been followed and if there are no obvious refactoring opportunities (there almost always are).
The LLM does most of the coding, yet I wouldn't call it "vibe coding" at all.
"Tele coding" would be more appropriate.
I’ve also found that a bigger focus on expanding my agents.md as the project rolls on has led to less headaches overall and more consistency (non-surprisingly). It’s the same as asking juniors to reflect on the work they’ve completed and to document important things that can help them in the future. Software Manger is a good way to put this.
AGENTS.md should mostly point to real documentation and design files that humans will also read and keep up to date. It's rare that something about a project is only of interest to AI agents.
I really like your analogy of LLMs as 'unreliable interns'. The shift from being a 'coder' to a 'software manager' who enforces documentation and grounding is the only way to scale these tools. Without an architecture.md or similar grounding, the context drift eventually makes the AI-generated code a liability rather than an asset. It's about moving the complexity from the syntax to the specification.
Oh no, maybe the V-Model was right all the time? And right sizing increments with control stops after them. No wonder these matrix multiplications start to behave like humans, that is what we wanted them to do.
So basically you’re saying LLMs are helping us be better humans?
Better humans? How and where?
It's nice to have it written down in a concise form. I shared it with my team as some engineers have been struggling with AI, and I think this (just trying to one-shot without planning) could be why.
It's alchemy all over again.
Alchemy involved a lot of do-it-yourself though. With AI it is like someone else does all the work (well, almost all the work).
It was mainly a jab at the protoscientific nature of it.
Reproducing experimental results across models and vendors is trivial and cheap nowadays.
Not if anthropic goes further in obfuscating the output of claude code.
> the workflow I’ve settled into is radically different from what most people do with AI coding tools
This looks exactly like what anthropic recommends as the best practice for using Claude Code. Textbook.
It also exposes a major downside of this approach: if you don't plan perfectly, you'll have to start over from scratch if anything goes wrong.
I've found a much better approach in doing a design -> plan -> execute in batches, where the plan is no more than 1,500 lines, used as a proxy for complexity.
My 30,000 LOC app has about 100,000 lines of plan behind it. Can't build something that big as a one-shot.
if you don't plan perfectly, you'll have to start over from scratch if anything goes wrong
This is my experience too, but it's pushed me to make much smaller plans and to commit things to a feature branch far more atomically so I can revert a step to the previous commit, or bin the entire feature by going back to main. I do this far more now than I ever did when I was writing the code by hand.
This is how developers should work regardless of how the code is being developed. I think this is a small but very real way AI has actually made me a better developer (unless I stop doing it when I don't use AI... not tried that yet.)
LLMs are really eager to start coding (as interns are eager to start working), so the sentence “don’t implement yet” has to be used very often at the beginning of any project.
Most LLM apps have a 'plan' or 'ask' mode for that.
Developers should work by wasting lots of time making the wrong thing?
I bet if they did a work and motion study on this approach they'd find the classic:
"Thinks they're more productive, AI has actually made them less productive"
But lots of lovely dopamine from this false progress that gets thrown away!
Developers should work by wasting lots of time making the wrong thing?
Yes. In fact, that's not emphatic enough: HELL YES!
More specifically, developers should experiment. They should test their hypothesis. They should try out ideas by designing a solution and creating a proof of concept, then throw that away and build a proper version based on what they learned.
If your approach to building something is to implement the first idea you have and move on then you are going to waste so much more time later refactoring things to fix architecture that paints you into corners, reimplementing things that didn't work for future use cases, fixing edge cases than you hadn't considered, and just paying off a mountain of tech debt.
I'd actually go so far as to say that if you aren't experimenting and throwing away solutions that don't quite work then you're only amassing tech debt and you're not really building anything that will last. If it does it's through luck rather than skill.
Also, this has nothing to do with AI. Developers should be working this way even if they handcraft their artisanal code carefully in vi.
> Developers should work by wasting lots of time making the wrong thing?
Yes? I can't even count how many times I worked on something my company deemed was valuable only for it to be deprecated or thrown away soon after. Or, how many times I solved a problem but apparently misunderstood the specs slightly and had to redo it. Or how many times we've had to refactor our code because scope increased. In fact, the very existence of the concepts of refactoring and tech debt proves that devs often spend a lot of time making the "wrong" thing.
Is it a waste? No, it solved the problem as understood at the time. And we learned stuff along the way.
Classic
https://metr.org/blog/2025-07-10-early-2025-ai-experienced-o...
> design -> plan -> execute in batches
This is the way for me as well. Have a high-level master design and plan, but break it apart into phases that are manageable. One-shotting anything beyond a todo list and expecting decent quality is still a pipe dream.
> if you don't plan perfectly, you'll have to start over from scratch if anything goes wrong.
You just revert what the AI agent changed and revise/iterate on the previous step - no need to start over. This can of course involve restricting the work to a smaller change so that the agent isn't overwhelmed by complexity.
100,000 lines is approx. one million words. The average person reads at 250wpm. The entire thing would take 66 hours just to read, assuming you were approaching it like a fiction book, not thinking anything over
wtf, why would you write 100k lines of plan to produce 30k loc.. JUST WRITE THE CODE!!!
They didn't write 100k plan lines. The llm did (99.9% of it at least or more). Writing 30k by hand would take weeks if not months. Llms do it in an afternoon.
Just reading that plan would take weeks or months
You don't start with 100k lines, you work in batches that are digestible. You read it once, then move on. The lines add up pretty quickly considering how fast Claude works. If you think about the difference in how many characters it takes to describe what code is doing in English, it's pretty reasonable.
And my weeks or months of work beats an LLMs 10/10 times. There are no shortcuts in life.
Might be true for you. But there are plenty of top tier engineers who love LLMs. So it works for some. Not for others.
And of course there are shortcuts in life. Any form of progress whether its cars, medicine, computers or the internet are all shortcuts in life. It makes life easier for a lot of people.
I have no doubts that it does for many people. But the time/cost tradeoff is still unquestionable. I know I could create what LLMs do for me in the frontend/backend in most cases as good or better - I know that, because I've done it at work for years. But to create a somewhat complex app with lots of pages/features/apis etc. would take me months if not a year++ since I'd be working on it only on the weekends for a few hours. Claude code helps me out by getting me to my goal in a fraction of the time. Its superpower lies not only in doign what I know but faster, but in doing what I don't know as well.
I yield similar benefits at work. I can wow management with LLM assited/vibe coded apps. What previously would've taken a multi-man team weeks of planning and executing, stand ups, jour fixes, architecture diagrams, etc. can now be done within a single week by myself. For the type of work I do, managers do not care whether I could do it better if I'd code it myself. They are amazed however that what has taken months previously, can be done in hours nowadays. And I for sure will try to reap benefits of LLMs for as long as they don't replace me rather than being idealistic and fighting against them.
That's not (or should not be what's happening).
They write a short high level plan (let's say 200 words). The plan asks the agent to write a more detailed implementation plan (written by the LLM, let's say 2000-5000 words).
They read this plan and adjust as needed, even sending it to the agent for re-dos.
Once the implementation plan is done, they ask the agent to write the actual code changes.
Then they review that and ask for fixes, adjustments, etc.
This can be comparable to writing the code yourself but also leaves a detailed trail of what was done and why, which I basically NEVER see in human generated code.
That alone is worth gold, by itself.
And on top of that, if you're using an unknown platform or stack, it's basically a rocket ship. You bootstrap much faster. Of course, stay on top of the architecture, do controlled changes, learn about the platform as you go, etc.
I take this concept and I meta-prompt it even more.
I have a road map (AI generated, of course) for a side project I'm toying around with to experiment with LLM-driven development. I read the road map and I understand and approve it. Then, using some skills I found on skills.sh and slightly modified, my workflow is as such:
1. Brainstorm the next slice
It suggests a few items from the road map that should be worked on, with some high level methodology to implement. It asks me what the scope ought to be and what invariants ought to be considered. I ask it what tradeoffs could be, why, and what it recommends, given the product constraints. I approve a given slice of work.
NB: this is the part I learn the most from. I ask it why X process would be better than Y process given the constraints and it either corrects itself or it explains why. "Why use an outbox pattern? What other patterns could we use and why aren't they the right fit?"
2. Generate slice
After I approve what to work on next, it generates a high level overview of the slice, including files touched, saved in a MD file that is persisted. I read through the slice, ensure that it is indeed working on what I expect it to be working on, and that it's not scope creeping or undermining scope, and I approve it. It then makes a plan based off of this.
3. Generate plan
It writes a rather lengthy plan, with discrete task bullets at the top. Beneath, each step has to-dos for the llm to follow, such as generating tests, running migrations, etc, with commit messages for each step. I glance through this for any potential red flags.
4. Execute
This part is self explanatory. It reads the plan and does its thing.
I've been extremely happy with this workflow. I'll probably write a blog post about it at some point.
Haven't seen a single useful thing produced by this garbage process you describe, I only see the people using it becoming useless themselves.
Keep frying your brain with neural network powered autocomplete.
> Haven't seen a single useful thing produced by this garbage process you describe
By using it first-hand or by a colleague? And useful to whom, you, or the person writing it? There are plenty of people in this thread who have actually used this "garbage process," myself included, to produce stuff we, and our colleagues, find is useful.
What genuinely new thing have you produced?
Dunno. My 80k+ LOC personal life planner, with a native android app, eink display view still one shots most features/bugs I encounter. I just open a new instance let it know what I want and 5min later it's done.
If you wouldn't mind sharing more about this in the future I'd love to read about it.
I've been thinking about doing something like that myself because I'm one of those people who have tried countless apps but there's always a couple deal breakers that cause me to drop the app.
I figured trying to agentically develop a planner app with the exact feature set I need would be an interesting and fun experiment.
In 5 min you are one shotting smaller changes to the larger code base right? Not the entire 80k likes which was the other comments point afaict.
Yeah, then I guess I misunderstood the post. Its smaller features one by one ofc.
Both can be true. I have personally experienced both.
Some problems AI surprised me immensely with fast, elegant efficient solutions and problem solving. I've also experienced AI doing totally absurd things that ended up taking multiple times longer than if I did it manually. Sometimes in the same project.
What is a personal life planner?
Todos, habits, goals, calendar, meals, notes, bookmarks, shopping lists, finances. More or less that with Google cal integration, garmin Integration (Auto updates workout habits, weight goals) family sharing/gamification, daily/weekly reviews, ai summaries and more. All built by just prompting Claude for feature after feature, with me writing 0 lines.
Ah, I imagined actual life planning as in asking AI what to do, I was morbidly curious.
Prompting basic notes apps is not as exciting but I can see how people who care about that also care about it being exactly a certain way, so I think get your excitement.
Is it on GH?
It was when I mvp'd it 3 weeks ago. Then I removed it as I was toying with the idea of somehow monetizing it. Then I added a few features which would make monetization impossible (e.g. How the app obtains etf/stock prices live and some other things). I reckon I could remove those and put in gh during the week if I don't forget. The quality of the Web app is SaaS grade IMO. Keyboard shortcuts, cmd+k, natural language parsing, great ui that doesn't look like made by ai in 5min. Might post here the link.
I do the same. I also cross-ask gemini and claude about the plan during iterations, sometimes make several separate plans.
> Notice the language: “deeply”, “in great details”, “intricacies”, “go through everything”. This isn’t fluff. Without these words, Claude will skim. It’ll read a file, see what a function does at the signature level, and move on. You need to signal that surface-level reading is not acceptable.
This makes no sense to my intuition of how an LLM works. It's not that I don't believe this works, but my mental model doesn't capture why asking the model to read the content "more deeply" will have any impact on whatever output the LLM generates.
It's the attention mechanism at work, along with a fair bit of Internet one-up-manship. The LLM has ingested all of the text on the Internet, as well as Github code repositories, pull requests, StackOverflow posts, code reviews, mailing lists, etc. In a number of those content sources, there will be people saying "Actually, if you go into the details of..." or "If you look at the intricacies of the problem" or "If you understood the problem deeply" followed by a very deep, expert-level explication of exactly what you should've done differently. You want the model to use the code in the correction, not the one in the original StackOverflow question.
Same reason that "Pretend you are an MIT professor" or "You are a leading Python expert" or similar works in prompts. It tells the model to pay attention to the part of the corpus that has those terms, weighting them more highly than all the other programming samples that it's run across.
I don’t think this is a result of the base training data („the internet“). It’s a post training behavior, created during reinforcement learning. Codex has a totally different behavior in that regard. Codex reads per default a lot of potentially relevant files before it goes and writes files.
Maybe you remember that, without reinforcement learning, the models of 2019 just completed the sentences you gave them. There were no tool calls like reading files. Tool calling behavior is company specific and highly tuned to their harnesses. How often they call a tool, is not part of the base training data.
Modern LLM are certainly fine tuned on data that includes examples of tool use, mostly the tools built into their respective harnesses, but also external/mock tools so they dont overfit on only using the toolset they expect to see in their harnesses.
IDK the current state, but I remember that, last year, the open source coding harnesses needed to provide exactly the tools that the LLM expected, or the error rate went through the roof. Some, like grok and gemini, only recently managed to make tool calls somewhat reliable.
Of course I can't be certain, but I think the "mixture of experts" design plays into it too. Metaphorically, there's a mid-level manager who looks at your prompt and tries to decide which experts it should be sent to. If he thinks you won't notice, he saves money by sending it to the undergraduate intern.
Just a theory.
Notice that MOE isn’t different experts for different types of problems. It’s per token and not really connect to problem type.
So if you send a python code then the first one in function can be one expert, second another expert and so on.
Can you back this up with documentation? I don't believe that this is the case.
Check out Unsloths REAP models, you can outright delete a few of the lesser used experts without the model going braindead since they all can handle each token but some are better posed to do so.
This is such a good explanation. Thanks
>> Same reason that "Pretend you are an MIT professor" or "You are a leading Python expert" or similar works in prompts.
This pretend-you-are-a-[persona] is cargo cult prompting at this point. The persona framing is just decoration.
A brief purpose statement describing what the skill [skill.md] does is more honest and just as effective.
You will never convince me that this isn't confirmation bias, or the equivalent of a slot machine player thinking the order in which they push buttons impacts the output, or some other gambler-esque superstition.
These tools are literally designed to make people behave like gamblers. And its working, except the house in this case takes the money you give them and lights it on fire.
Your ignorance is my opportunity. May I ask which markets you are developing for?
"The equivalent of saying, which slot machine were you sitting at It'll make me money"
GPS denied navigation. Good luck using llms to slop produce your way into competing with us lol.
That’s because it’s superstition.
Unless someone can come up with some kind of rigorous statistics on what the effect of this kind of priming is it seems no better than claiming that sacrificing your first born will please the sun god into giving us a bountiful harvest next year.
Sure, maybe this supposed deity really is this insecure and needs a jolly good pep talk every time he wakes up. or maybe you’re just suffering from magical thinking that your incantations had any effect on the random variable word machine.
The thing is, you could actually prove it, it’s an optimization problem, you have a model, you can generate the statistics, but no one as far as I can tell has been terribly forthcoming with that , either because those that have tried have decided to try to keep their magic spells secret, or because it doesn’t really work.
If it did work, well, the oldest trick in computer science is writing compilers, i suppose we will just have to write an English to pedantry compiler.
> If it did work, well, the oldest trick in computer science is writing compilers, i suppose we will just have to write an English to pedantry compiler.
"Add tests to this function" for GPT-3.5-era models was much less effective than "you are a senior engineer. add tests for this function. as a good engineer, you should follow the patterns used in these other three function+test examples, using this framework and mocking lib." In today's tools, "add tests to this function" results in a bunch of initial steps to look in common places to see if that additional context already exists, and then pull it in based on what it finds. You can see it in the output the tools spit out while "thinking."
So I'm 90% sure this is already happening on some level.
I actually have a prompt optimizer skill that does exactly this.
https://github.com/solatis/claude-config
It’s based entirely off academic research, and a LOT of research has been done in this area.
One of the papers you may be interested in is “emotion prompting”, eg “it is super important for me that you do X” etc actually works.
“Large Language Models Understand and Can be Enhanced by Emotional Stimuli”
https://arxiv.org/abs/2307.11760
I think "understand this directory deeply" just gives more focus for the instruction. So it's like "burn more tokens for this phase than you normally would".
i suppose we will just have to write an English to pedantry compiler.
A common technique is to prompt in your chosen AI to write a longer prompt to get it to do what you want. It's used a lot in image generation. This is called 'prompt enhancing'.
> That’s because it’s superstition.
This field is full of it. Practices are promoted by those who tie their personal or commercial brand to it for increased exposure, and adopted by those who are easily influenced and don't bother verifying if they actually work.
This is why we see a new Markdown format every week, "skills", "benchmarks", and other useless ideas, practices, and measurements. Consider just how many "how I use AI" articles are created and promoted. Most of the field runs on anecdata.
It's not until someone actually takes the time to evaluate some of these memes, that they find little to no practical value in them.[1]
[1]: https://news.ycombinator.com/item?id=47034087
Its a wild time to be in software development. Nobody(1) actually knows what causes LLMs to do certain things, we just pray the prompt moves the probabilities the right way enough such that it mostly does what we want. This used to be a field that prided itself on deterministic behavior and reproducibility.
Now? We have AGENTS.md files that look like a parent talking to a child with all the bold all-caps, double emphasis, just praying that's enough to be sure they run the commands you want them to be running
(1 Outside of some core ML developers at the big model companies)
It’s like playing a fretless instrument to me.
Practice playing songs by ear and after 2 weeks, my brain has developed an inference model of where my fingers should go to hit any given pitch.
Do I have any idea how my brain’s model works? No! But it tickles a different part of my brain and I like it.
Sufficiently advanced technology has become like magic: you have to prompt the electronic genie with the right words or it will twist your wishes.
Light some incense, and you too can be a dystopian space tech support, today! Praise Omnissiah!
are we the orks?
For Claude at least, the more recent guidance from Anthropic is to not yell at it. Just clear, calm, and concise instructions.
Yep, with Claude saying "please" and "thank you" actually works. If you build rapport with Claude, you get rewarded with intuition and creativity. Codex, on the other hand, you have to slap it around like a slave gollum and it will do exactly what you tell it to do, no more, no less.
this is psychotic why is this how this works lol
Speculation only obviously: highly-charged conversations cause the discussion to be channelled to general human mitigation techniques and for the 'thinking agent' to be diverted to continuations from text concerned with the general human emotional experience.
Sometimes I daydream about people screaming at their LLM as if it was a TV they were playing video games on.
wait seriously? lmfao
thats hilarious. i definitely treat claude like shit and ive noticed the falloff in results.
if there's a source for that i'd love to read about it.
I don't have a source offhand, but I think it may have been part of the 4.5 release? Older models definitely needed caps and words like critical, important, never, etc... but Anthropic published something that said don't do that anymore.
If you think about where in the training data there is positivity vs negativity it really becomes equivalent to having a positive or negative mindset regarding a standing and outcome in life.
For awhile(maybe a year ago?) it seemed like verbal abuse was the best way to make Claude pay attention. In my head, it was impacting how important it deemed the instruction. And it definitely did seem that way.
i make claude grovel at my feet and tell me in detail why my code is better than its code
Consciousness is off the table but they absolutely respond to environmental stimulus and vibes.
See, uhhh, https://pmc.ncbi.nlm.nih.gov/articles/PMC8052213/ and maybe have a shot at running claude while playing Enya albums on loop.
/s (??)
i have like the faintest vague thread of "maybe this actually checks out" in a way that has shit all to do with consciousness
sometimes internet arguments get messy, people die on their hills and double / triple down on internet message boards. since historic internet data composes a bit of what goes into an llm, would it make sense that bad-juju prompting sends it to some dark corners of its training model if implementations don't properly sanitize certain negative words/phrases ?
in some ways llm stuff is a very odd mirror that haphazardly regurgitates things resulting from the many shades of gray we find in human qualities.... but presents results as matter of fact. the amount of internet posts with possible code solutions and more where people egotistically die on their respective hills that have made it into these models is probably off the charts, even if the original content was a far cry from a sensible solution.
all in all llm's really do introduce quite a bit of a black box. lot of benefits, but a ton of unknowns and one must be hyperviligant to the possible pitfalls of these things... but more importantly be self aware enough to understand the possible pitfalls that these things introduce to the person using them. they really possibly dangerously capitalize on everyones innate need to want to be a valued contributor. it's really common now to see so many people biting off more than they can chew, often times lacking the foundations that would've normally had a competent engineer pumping the brakes. i have a lot of respect/appreciation for people who might be doing a bit of claude here and there but are flat out forward about it in their readme and very plainly state to not have any high expectations because _they_ are aware of the risks involved here. i also want to commend everyone who writes their own damn readme.md.
these things are for better or for worse great at causing people to barrel forward through 'problem solving', which is presenting quite a bit of gray area on whether or not the problem is actually solved / how can you be sure / do you understand how the fix/solution/implementation works (in many cases, no). this is why exceptional software engineers can use this technology insanely proficiently as a supplementary worker of sorts but others find themselves in a design/architect seat for the first time and call tons of terrible shots throughout the course of what it is they are building. i'd at least like to call out that people who feel like they "can do everything on their own and don't need to rely on anyone" anymore seem to have lost the plot entirely. there are facets of that statement that might be true, but less collaboration especially in organizations is quite frankly the first steps some people take towards becoming delusional. and that is always a really sad state of affairs to watch unfold. doing stuff in a vaccuum is fun on your own time, but forcing others to just accept things you built in a vaccuum when you're in any sort of team structure is insanely immature and honestly very destructive/risky. i would like to think absolutely no one here is surprised that some sub-orgs at Microsoft force people to use copilot or be fired, very dangerous path they tread there as they bodyslam into place solutions that are not well understood. suddenly all the leadership decisions at many companies that have made to once again bring back a before-times era of offshoring work makes sense: they think with these technologies existing the subordinate culture of overseas workers combined with these techs will deliver solutions no one can push back on. great savings and also no one will say no.
How anybody can read stuff like this and still take all this seriously is beyond me. This is becoming the engineering equivalent of astrology.
The evolution of software engineering is fascinating to me. We started by coding in thin wrappers over machine code and then moved on to higher-level abstractions. Now, we've reached the point where we discuss how we should talk to a mystical genie in a box.
I'm not being sarcastic. This is absolutely incredible.
Anthropic recommends doing magic invocations: https://simonwillison.net/2025/Apr/19/claude-code-best-pract...
It's easy to know why they work. The magic invocation increases test-time compute (easy to verify yourself - try!). And an increase in test-time compute is demonstrated to increase answer correctness (see any benchmark).
It might surprise you to know that the only different between GPT 5.2-low and GPT 5.2-xhigh is one of these magic invocations. But that's not supposed to be public knowledge.
I think this was more of a thing on older models. Since I started using Opus 4.5 I have not felt the need to do this.
Feel free to run your own tests and see if the magic phrases do or do not influence the output. Have it make a Todo webapp with and without those phrases and see what happens!
That's not how it works. It's not on everyone else to prove claims false, it's on you (or the people who argue any of this had a measurable impact) to prove it actually works. I've seen a bunch of articles like this, and more comments. Nobody I've ever seen has produced any kind of measurable metrics of quality based on one approach vs another. It's all just vibes.
Without something quantifiable it's not much better then someone who always wears the same jersey when their favorite team plays, and swears they play better because of it.
These coding agents are literally Language Models. The way you structure your prompting language affect the actual output.
If you read the transformer paper, or get any book on NLP, you will see that this is not magic incantation; it's purely the attention mechanism at work. Or you can just ask Gemini or Claude why these prompts work.
But I get the impression from your comment that you have a fixed idea, and you're not really interested in understanding how or why it works.
If you think like a hammer, everything will look like a nail.
I know why it works, to varying and unmeasurable degrees of success. Just like if I poke a bull with a sharp stick, I know it's gonna get it's attention. It might choose to run away from me in one of any number of directions, or it might decide to turn around and gore me to death. I can't answer that question with any certainty then you can.
The system is inherently non-deterministic. Just because you can guide it a bit, doesn't mean you can predict outcomes.
> The system is inherently non-deterministic.
The system isn't randomly non-deterministic; it is statistically probabilistic.
The next-token prediction and the attention mechanism is actually a rigorous deterministic mathematical process. The variation in output comes from how we sample from that curve, and the temperature used to calibrate the model. Because the underlying probabilities are mathematically calculated, the system's behavior remains highly predictable within statistical bounds.
Yes, it's a departure from the fully deterministic systems we're used to. But that's not different than the many real world systems: weather, biology, robotics, quantum mechanics. Even the computer you're reading this right now is full of probabilistic processes, abstracted away through sigmoid-like functions that push the extremes to 0s and 1s.
A lot of words to say that for all intents and purposes... it's nondeterministic.
> Yes, it's a departure from the fully deterministic systems we're used to.
A system either produces the same output given the same input[1], or doesn't.
LLMs are nondeterministic by design. Sure, you can configure them with a zero temperature, a static seed, and so on, but they're of no use to anyone in that configuration. The nondeterminism is what gives them the illusion of "creativity", and other useful properties.
Classical computers, compilers, and programming languages are deterministic by design, even if they do contain complex logic that may affect their output in unpredictable ways. There's a world of difference.
[1]: Barring misbehavior due to malfunction, corruption or freak events of nature (cosmic rays, etc.).
But we can predict the outcomes, though. That's what we're saying, and it's true. Maybe not 100% of the time, but maybe it helps a significant amount of the time and that's what matters.
Is it engineering? Maybe not. But neither is knowing how to talk to junior developers so they're productive and don't feel bad. The engineering is at other levels.
> But we can predict the outcomes [...] Maybe not 100% of the time
So 60% of the time, it works every time.
... This fucking industry.
Do you actively use LLMs to do semi-complex coding work? Because if not, it will sound mumbo-jumbo to you. Everyone else can nod along and read on, as they’ve experienced all of it first hand.
You've missed the point. This isn't engineering, it's gambling.
You could take the exact same documents, prompts, and whatever other bullshit, run it on the exact same agent backed by the exact same model, and get different results every single time. Just like you can roll dice the exact same way on the exact same table and you'll get two totally different results. People are doing their best to constrain that behavior by layering stuff on top, but the foundational tech is flawed (or at least ill suited for this use case).
That's not to say that AI isn't helpful. It certainly is. But when you are basically begging your tools to please do what you want with magic incantations, we've lost the fucking plot somewhere.
I think that's a pretty bold claim, that it'd be different every time. I'd think the output would converge on a small set of functionally equivalent designs, given sufficiently rigorous requirements.
And even a human engineer might not solve a problem the same way twice in a row, based on changes in recent inspirations or tech obsessions. What's the difference, as long as it passes review and does the job?
> You could take the exact same documents, prompts, and whatever other bullshit, run it on the exact same agent backed by the exact same model, and get different results every single time
This is more of an implementation detail/done this way to get better results. A neural network with fixed weights (and deterministic floating point operations) returning a probability distribution, where you use a pseudorandom generator with a fixed seed called recursively will always return the same output for the same input.
these sort-of-lies might help:
think of the latent space inside the model like a topological map, and when you give it a prompt, you're dropping a ball at a certain point above the ground, and gravity pulls it along the surface until it settles.
caveat though, thats nice per-token, but the signal gets messed up by picking a token from a distribution, so each token you're regenerating and re-distorting the signal. leaning on language that places that ball deep in a region that you want to be makes it less likely that those distortions will kick it out of the basin or valley you may want to end up in.
if the response you get is 1000 tokens long, the initial trajectory needed to survive 1000 probabilistic filters to get there.
or maybe none of that is right lol but thinking that it is has worked for me, which has been good enough
Hah! Reading this, my mind inverted it a bit, and I realized ... it's like the claw machine theory of gradient descent. Do you drop the claw into the deepest part of the pile, or where there's the thinnest layer, the best chance of grabbing something specific? Everyone in everu bar has a theory about claw machines. But the really funny thing that unites LLMs with claw machines is that the biggest question is always whether they dropped the ball on purpose.
The claw machine is also a sort-of-lie, of course. Its main appeal is that it offers the illusion of control. As a former designer and coder of online slot machines... totally spin off into pages on this analogy, about how that illusion gets you to keep pulling the lever... but the geographic rendition you gave is sort of priceless when you start making the comparison.
My mental model for them is plinko boards. Your prompt changes the spacing between the nails to increase the probability in certain directions as your chip falls down.
i literally suggested this metaphor earlier yesterday to someone trying to get agents to do stuff they wanted, that they had to set up their guardrails in a way that you can let the agents do what they're good at, and you'll get better results because you're not sitting there looking at them.
i think probably once you start seeing that the behavior falls right out of the geometry, you just start looking at stuff like that. still funny though.
Its very logical and pretty obvious when you do code generation. If you ask the same model, to generate code by starting with:
- You are a Python Developer... or - You are a Professional Python Developer... or - You are one of the World most renowned Python Experts, with several books written on the subject, and 15 years of experience in creating highly reliable production quality code...
You will notice a clear improvement in the quality of the generated artifacts.
Do you think that Anthropic don’t include things like this in their harness / system prompts? I feel like this kind of prompts are uneccessary with Opus 4.5 onwards, obviously based on my own experience (I used to do this, on switching to opus I stopped and have implemented more complex problems, more successfully).
I am having the most success describing what I want as humanly as possible, describing outcomes clearly, making sure the plan is good and clearing context before implementing.
My colleague swears by his DHH claude skill https://danieltenner.com/dhh-is-immortal-and-costs-200-m/
That's different. You are pulling the model, semantically, closer to the problem domain you want it to attack.
That's very different from "think deeper". I'm just curious about this case in specific :)
I don't know about some of those "incantations", but it's pretty clear that an LLM can respond to "generate twenty sentences" vs. "generate one word". That means you can indeed coax it into more verbosity ("in great detail"), and that can help align the output by having more relevant context (inserting irrelevant context or something entirely improbable into LLM output and forcing it to continue from there makes it clear how detrimental that can be).
Of course, that doesn't mean it'll definitely be better, but if you're making an LLM chain it seems prudent to preserve whatever info you can at each step.
If I say “you are our domain expert for X, plan this task out in great detail” to a human engineer when delegating a task, 9 times out of 10 they will do a more thorough job. It’s not that this is voodoo that unlocks some secret part of their brain. It simply establishes my expectations and they act accordingly.
To the extent that LLMs mimic human behaviour, it shouldn’t be a surprise that setting clear expectations works there too.
The LLM will do what you ask it to unless you don't get nuanced about it. Myself and others have noticed that LLM's work better when your codebase is not full of code smells like massive godclass files, if your codebase is discrete and broken up in a way that makes sense, and fits in your head, it will fit in the models head.
Maybe the training data that included the words like "skim" also provided shallower analysis than training that was close to the words "in great detail", so the LLM is just reproducing those respective words distribution when prompted with directions to do either.
Apparently LLM quality is sensitive to emotional stimuli?
"Large Language Models Understand and Can be Enhanced by Emotional Stimuli": https://arxiv.org/abs/2307.11760
It is as the author said, it'll skim the content unless otherwise prompted to do so. It can read partial file fragments; it can emit commands to search for patterns in the files. As opposed to carefully reading each file and reasoning through the implementation. By asking it to go through in detail you are telling it to not take shortcuts and actually read the actual code in full.
The original “chain of thought” breakthrough was literally to insert words like “Wait” and “Let’s think step by step”.
It’s actually really common. If you look at Claude Code’s own system prompts written by Anthropic, they’re littered with “CRITICAL (RULE 0):” type of statements, and other similar prompting styles.
Where can I find those?
This analysis is a good starting point: https://southbridge-research.notion.site/Prompt-Engineering-...
My guess would be that there’s a greater absolute magnitude of the vectors to get to the same point in the knowledge model.
The disconnect might be that there is a separation between "generating the final answer for the user" and "researching/thinking to get information needed for that answer". Saying "deeply" prompts it to read more of the file (as in, actually use the `read` tool to grab more parts of the file into context), and generate more "thinking" tokens (as in, tokens that are not shown to the user but that the model writes to refine its thoughts and improve the quality of its answer).
The author is referring to how the framing of your prompt informs the attention mechanism. You are essentially hinting to the attention mechanism that the function's implementation details have important context as well.
Yeah, it's definitely a strange new world we're in, where I have to "trick" the computer into cooperating. The other day I told Claude "Yes you can", and it went off and did something it just said it couldn't do!
Solid dad move. XD
Is parenting making us better at prompt engineering, or is it the other way around?
Better yet, I have Codex, Gemini, and Claude as my kids, running around in my code playground. How do I be a good parent and not play favorites?
We all know Gemini is your artsy, Claude is your smartypants, and Codex is your nerd.
You bumped the token predictor into the latent space where it knew what it was doing : )
The little language model that could.
—HAL, open the shuttle bay doors.
(chirp)
—HAL, please open the shuttle bay doors.
(pause)
—HAL!
—I'm afraid I can't do that, Dave.
It's very much believable, to me.
In image generation, it's fairly common to add "masterpiece", for example.
I don't think of the LLM as a smart assistant that knows what I want. When I tell it to write some code, how does it know I want it to write the code like a world renowned expert would, rather than a junior dev?
I mean, certainly Anthropic has tried hard to make the former the case, but the Titanic inertia from internet scale data bias is hard to overcome. You can help the model with these hints.
Anyway, luckily this is something you can empirically verify. This way, you don't have to take anyone's word. If anything, if you find I'm wrong in your experiments, please share it!
Its effectiveness is even more apparent with older smaller LLMs, people who interact with LLMs now never tried to wrangle llama2-13b into pretending to be a dungeon master...
if it’s so smart, why do i need to learn to use it?
Strings of tokens are vectors. Vectors are directions. When you use a phrase like that you are orienting the vector of the overall prompt toward the direction of depth, in its map of conceptual space.
One of the well defined failure modes for AI agents/models is "laziness." Yes, models can be "lazy" and that is an actual term used when reviewing them.
I am not sure if we know why really, but they are that way and you need to explicitly prompt around it.
I've encountered this failure mode, and the opposite of it: thinking too much. A behaviour I've come to see as some sort of pseudo-neuroticism.
Lazy thinking makes LLMs do surface analysis and then produce things that are wrong. Neurotic thinking will see them over-analyze, and then repeatedly second-guess themselves, repeatedly re-derive conclusions.
Something very similar to an anxiety loop in humans, where problems without solutions are obsessed about in circles.
yeah i experienced this the other day when asking claude code to build an http proxy using an afsk modem software to communicate over the computers sound card. it had an absolute fit tuning the system and would loop for hours trying and doubling back. eventually after some change in prompt direction to think more deeply and test more comprehensively it figured it out. i certainly had no idea how to build a afsk modem.
I don’t use plan.md docs either, but I recognise the underlying idea: you need a way to keep agent output constrained by reality.
My workflow is more like scaffold -> thin vertical slices -> machine-checkable semantics -> repeat.
Concrete example: I built and shipped a live ticketing system for my club (Kolibri Tickets). It’s not a toy: real payments (Stripe), email delivery, ticket verification at the door, frontend + backend, migrations, idempotency edges, etc. It’s running and taking money.
The reason this works with AI isn’t that the model “codes fast”. It’s that the workflow moves the bottleneck from “typing” to “verification”, and then engineers the verification loop:
If you run it open-loop (prompt -> giant diff -> read/debug), you get the “illusion of velocity” people complain about. If you run it closed-loop (scaffold + constraints + verifiers), you can actually ship faster because you’re not paying the integration cost repeatedly.Plan docs are one way to create shared state and prevent drift. A runnable scaffold + verification harness is another.
Now that code is cheap, I ensured my side project has unit/integration tests (will enforce 100% coverage), Playwright tests, static typing (its in Python), scripts for all tasks. Will learn mutation testing too (yes, its overkill). Now my agent works upto 1 hour in loops and emits concise code I dont have to edit much.
I actually don't really like a few of things about this approach.
First, the "big bang" write it all at once. You are going to end up with thousands of lines of code that were monolithically produced. I think it is much better to have it write the plan and formulate it as sensible technical steps that can be completed one at a time. Then you can work through them. I get that this is not very "vibe"ish but that is kind of the point. I want the AI to help me get to the same point I would be at with produced code AND understanding of it, just accelerate that process. I'm not really interested in just generating thousands of lines of code that nobody understands.
Second, the author keeps refering to adjusting the behaviour, but never incorporating that into long lived guidance. To me, integral with the planning process is building an overarching knowledge base. Every time you're telling it there's something wrong, you need to tell it to update the knowledge base about why so it doesn't do it again.
Finally, no mention of tests? Just quick checks? To me, you have to end up with comprehensive tests. Maybe to the author it goes without saying, but I find it is integral to build this into the planning. Certain stages you will want certain types of tests. Some times in advance of the code (so TDD style) other times built alongside it or after.
It's definitely going to be interesting to see how software methodology evolves to incorporate AI support and where it ultimately lands.
The articles approach matches mine, but I've learned from exactly the things you're pointing out.
I get the PLAN.md (or equivalent) to be separated into "phases" or stages, then carefully prompt (because Claude and Codex both love to "keep going") it to only implement that stage, and update the PLAN.md
Tests are crucial too, and form another part of the plan really. Though my current workflow begins to build them later in the process than I would prefer...
I use Claude Code for lecture prep.
I craft a detailed and ordered set of lecture notes in a Quarto file and then have a dedicated claude code skill for translating those notes into Slidev slides, in the style that I like.
Once that's done, much like the author, I go through the slides and make commented annotations like "this should be broken into two slides" or "this should be a side-by-side" or "use your generate clipart skill to throw an image here alongside these bullets" and "pull in the code example from ../examples/foo." It works brilliantly.
And then I do one final pass of tweaking after that's done.
But yeah, annotations are super powerful. Token distance in-context and all that jazz.
Can I ask how you annotate the feedback for it? Just with inline comments like `# This should be changed to X`?
The author mentions annotations but doesn't go into detail about how to feed the annotations to Claude.
Slidev is markdown, so i do it in html comments. Usually something like:
or And then, when I finish annotating I just say: "Address all the TODOCLAUDEs"is your skill open source
Not yet... but also I'm not sure it makes a lot of sense to be open source. It's super specific to how I like to build slide decks and to my personal lecture style.
But it's not hard to build one. The key for me was describing, in great detail:
1. How I want it to read the source material (e.g., H1 means new section, H2 means at least one slide, a link to an example means I want code in the slide)
2. How to connect material to layouts (e.g., "comparison between two ideas should be a two-cols-title," "walkthrough of code should be two-cols with code on right," "learning objectives should be side-title align:left," "recall should be side-title align:right")
Then the workflow is:
1. Give all those details and have it do a first pass.
2. Give tons of feedback.
3. At the end of the session, ask it to "make a skill."
4. Manually edit the skill so that you're happy with the examples.
> Read deeply, write a plan, annotate the plan until it’s right, then let Claude execute the whole thing without stopping, checking types along the way.
As others have already noted, this workflow is exactly what the Google Antigravity agent (based off Visual Studio Code) has been created for. Antigravity even includes specialized UI for a user to annotate selected portions of an LLM-generated plan before iterating it.
One significant downside to Antigravity I have found so far is the fact that even though it will properly infer a certain technical requirement and clearly note it in the plan it generates (for example, "this business reporting column needs to use a weighted average"), it will sometimes quietly downgrade such a specialized requirement (for example, to a non-weighted average), without even creating an appropriate "WARNING:" comment in the generated code. Especially so when the relevant codebase already includes a similar, but not exactly appropriate API. My repetitive prompts to ALWAYS ask about ANY implementation ambiguities WHATSOEVER go unanswered.
From what I gather Claude Code seems to be better than other agents at always remembering to query the user about implementation ambiguities, so maybe I will give Claude Code a shot over Antigravity.
This all looks fine for someone who can't code, but for anyone with even a moderate amount of experience as a developer all this planning and checking and prompting and orchestrating is far more work than just writing the code yourself.
There's no winner for "least amount of code written regardless of productivity outcomes.", except for maybe Anthropic's bank account.
I really don't understand why there are so many comments like this.
Yesterday I had Claude write an audit logging feature to track all changes made to entities in my app. Yeah you get this for free with many frameworks, but my company's custom setup doesn't have it.
It took maybe 5-10 minutes of wall-time to come up with a good plan, and then ~20-30 min for Claude implement, test, etc.
That would've taken me at least a day, maybe two. I had 4-5 other tasks going on in other tabs while I waited the 20-30 min for Claude to generate the feature.
After Claude generated, I needed to manually test that it worked, and it did. I then needed to review the code before making a PR. In all, maybe 30-45 minutes of my actual time to add a small feature.
All I can really say is... are you sure you're using it right? Have you _really_ invested time into learning how to use AI tools?
Same here. I did bounce off these tools a year ago. They just didn't work for me 60% of the time. I learned a bit in that initial experience though and walked away with some tasks ChatGPT could replace in my workflow. Mainly replacing scripts and reviewing single files or functions.
Fast forward to today and I tried the tools again--specifically Claude Code--about a week ago. I'm blown away. I've reproduced some tools that took me weeks at full-time roles in a single day. This is while reviewing every line of code. The output is more or less what I'd be writing as a principal engineer.
> The output is more or less what I'd be writing as a principal engineer.
I certainly hope this is not true, because then you're not competent for that role. Claude Code writes an absolutely incredible amount of unecessary and superfluous comments, it's makes asinine mistakes like forgetting to update logic in multiple places. It'll gladly drop the entire database when changing column formats, just as an example.
Trust me I'm very impressed at the progress AI has made, and maybe we'll get to the point where everything is 100% correct all the time and better than any human could write. I'm skeptical we can get there with the LLM approach though.
The problem is LLMs are great at simple implementation, even large amounts of simple implementation, but I've never seen it develop something more than trivial correctly. The larger problem is it's very often subtly but hugely wrong. It makes bad architecture decisions, it breaks things in pursuit of fixing or implementing other things. You can tell it has no concept of the "right" way to implement something. It very obviously lacks the "senior developer insight".
Maybe you can resolve some of these with large amounts of planning or specs, but that's the point of my original comment - at what point is it easier/faster/better to just write the code yourself? You don't get a prize for writing the least amount of code when you're just writing specs instead.
This is exactly what the article is about. The tradeoff is that you have to throughly review the plans and iterate on them, which is tiring. But the LLM will write good code faster than you, if you tell it what good code is.
My experience has so far been similar to the root commenter - at the stage where you need to have a long cycle with planning it's just slower than doing the writing + theory building on my own.
It's an okay mental energy saver for simpler things, but for me the self review in an actual production code context is much more draining than writing is.
I guess we're seeing the split of people for whom reviewing is easy and writing is difficult and vice versa.
Exactly; the original commenter seems determined to write-off AI as "just not as good as me".
The original article is, to me, seemingly not that novel. Not because it's a trite example, but because I've begun to experience massive gains from following the same basic premise as the article. And I can't believe there's others who aren't using like this.
I iterate the plan until it's seemingly deterministic, then I strip the plan of implementation, and re-write it following a TDD approach. Then I read all specs, and generate all the code to red->green the tests.
If this commenter is too good for that, then it's that attitude that'll keep him stuck. I already feel like my projects backlog is achievable, this year.
Strongly agree about the deterministic part. Even more important than a good design, the plan must not show any doubt, whether it's in the form of open questions or weasel words. 95% of the time those vague words mean I didn't think something through, and it will do something hideous in order to make the plan work
Several months ago, just for fun, I asked Claude (the web site, not Claude Code) to build a web page with a little animated cannon that shoots at the mouse cursor with a ballistic trajectory. It built the page in seconds, but the aim was incorrect; it always shot too low. I told it the aim was off. It still got it wrong. I prompted it several times to try to correct it, but it never got it right. In fact, the web page started to break and Claude was introducing nasty bugs.
More recently, I tried the same experiment, again with Claude. I used the exact same prompt. This time, the aim was exactly correct. Instead of spending my time trying to correct it, I was able to ask it to add features. I've spent more time writing this comment on HN than I spent optimizing this toy. https://claude.ai/public/artifacts/d7f1c13c-2423-4f03-9fc4-8...
My point is that AI-assisted coding has improved dramatically in the past few months. I don't know whether it can reason deeply about things, but it can certainly imitate a human who reasons deeply. I've never seen any technology improve at this rate.
> but I've never seen it develop something more than trivial correctly.
What are you working on? I personally haven't seen LLMs struggle with any kind of problem in months. Legacy codebase with great complexity and performance-critical code. No issue whatsoever regardless of the size of the task.
>I've never seen it develop something more than trivial correctly.
This is 100% incorrect, but the real issue is that the people who are using these llms for non-trivial work tend to be extremely secretive about it.
For example, I view my use of LLMs to be a competitive advantage and I will hold on to this for as long as possible.
The key part of my comment is "correctly".
Does it write maintainable code? Does it write extensible code? Does it write secure code? Does it write performant code?
My experience has been it failing most of these. The code might "work", but it's not good for anything more than trivial, well defined functions (that probably appeared in it's training data written by humans). LLMs have a fundamental lack of understanding of what they're doing, and it's obvious when you look at the finer points of the outcomes.
That said, I'm sure you could write detailed enough specs and provide enough examples to resolve these issues, but that's the point of my original comment - if you're just writing specs instead of code you're not gaining anything.
I find “maintainable code” the hardest bias to let go of. 15+ years of coding and design patterns are hard to let go.
But the aha moment for me was what’s maintainable by AI vs by me by hand are on different realms. So maintainable has to evolve from good human design patterns to good AI patterns.
Specs are worth it IMO. Not because if I can spec, I could’ve coded anyway. But because I gain all the insight and capabilities of AI, while minimizing the gotchas and edge failures.
> But the aha moment for me was what’s maintainable by AI vs by me by hand are on different realms
I don't find that LLMs are any more likely than humans to remember to update all of the places it wrote redundant functions. Generally far less likely, actually. So forgive me for treating this claim with a massive grain of salt.
> But the aha moment for me was what’s maintainable by AI vs by me by hand are on different realms. So maintainable has to evolve from good human design patterns to good AI patterns.
How do you square that with the idea that all the code still has to be reviewed by humans? Yourself, and your coworkers
I picture like semi conductors; the 5nm process is so absurdly complex that operators can't just peek into the system easily. I imagine I'm just so used to hand crafting code that I can't imagine not being able to peek in.
So maybe it's that we won't be reviewing by hand anymore? I.e. it's LLMs all the way down. Trying to embrace that style of development lately as unnatural as it feels. We're obv not 100% there yet but Claude Opus is a significant step in that direction and they keep getting better and better.
Then who is responsible when (not if) that code does horrible things? We have humans to blame right now. I just don’t see it happening personally because liability and responsibility are too important
For some software, sure but not most.
And you don’t blame humans anyways lol. Everywhere I’ve worked has had “blameless” postmortems. You don’t remove human review unless you have reasonable alternatives like high test coverage and other automated reviews.
We still have performance reviews and are fired. There’s a human that is responsible.
“It’s AI all the way down” is either nonsense on its face, or the industry is dead already.
To answer all of your questions:
yes, if I steer it properly.
It's very good at spotting design patterns, and implementing them. It doesn't always know where or how to implement them, but that's my job.
The specs and syntactic sugar are just nice quality of life benefits.
You’d be building blocks which compound over time. That’s been my experience anyway.
The compounding is much greater than my brain can do on its own.
> Yesterday I had Claude write an audit logging feature to track all changes made to entities in my app. Yeah you get this for free with many frameworks, but my company's custom setup doesn't have it.
But did you truly think about such feature? Like guarantees that it should follow (like how do it should cope with entities migration like adding a new field) or what the cost of maintaining it further down the line. This looks suspiciously like drive-by PR made on open-source projects.
> That would've taken me at least a day, maybe two.
I think those two days would have been filled with research, comparing alternatives, questions like "can we extract this feature from framework X?", discussing ownership and sharing knowledge,.. Jumping on coding was done before LLMs, but it usually hurts the long term viability of the project.
Adding code to a project can be done quite fast (hackatons,...), ensuring quality is what slows things down in any any well functioning team.
I mean, all I can really say is... if writing some logging takes you one or two days, are you sure you _really_ know how to code?
Ever worked on a distributed system with hundreds of millions of customers and seemingly endless business requirements?
Some things are complex.
You're right, you're better than me!
You could've been curious and ask why it would take 1-2 days, and I would've happily told you.
I'll bite, because it does seem like something that should be quick in a well-architected codebase. What was the situation? Was there something in this codebase that was especially suited to AI-development? Large amounts of duplication perhaps?
It's not particularly interesting.
I wanted to add audit logging for all endpoints we call, all places we call the DB, etc. across areas I haven't touched before. It would have taken me a while to track down all of the touchpoints.
Granted, I am not 100% certain that Claude didn't miss anything. I feel fairly confident that it is correct given that I had it research upfront, had multiple agents review, and it made the correct changes in the areas that I knew.
Also I'm realizing I didn't mention it included an API + UI for viewing events w/ pretty deltas
Well someone who says logging is easy never knows the difficulty of deciding "what" to log. And audit log is different beast altogether than normal logging
Audit logging is different than developer logging… companies will have entire teams dedicated to audit systems.
We're not as good at coding as you, naturally.
I'd find it deeply funny if the optimal vibe coding workflow continues to evolve to include more and more human oversight, and less and less agent autonomy, to the point where eventually someone makes a final breakthrough that they can save time by bypassing the LLM entirely and writing the code themselves. (Finally coming full circle.)
You mean there will be an invention to edit files directly instead of giving the specific code and location you want it to be written into the prompt?
Researching and planning a project is a generally usefully thing. This is something I've been doing for years, and have always had great results compared to just jumping in and coding. It makes perfect sense that this transfers to LLM use.
Well it's less mental load. It's like Tesla's FSD. Am I a better driver than the FSD? For sure. But is it nice to just sit back and let it drive for a bit even if it's suboptimal and gets me there 10% slower, and maybe slightly pisses off the guy behind me? Yes, nice enough to shell out $99/mo. Code implementation takes a toll on you in the same way that driving does.
I think the method in TFA is overall less stressful for the dev. And you can always fix it up manually in the end; AI coding vs manual coding is not either-or.
Since Opus 4.5, things have changed quite a lot. I find LLMs very useful for discussing new features or ideas, and Sonnet is great for executing your plan while you grab a coffee.
Most of these AI coding articles seem to be about greenfield development.
That said, if you're on a serious team writing professional software there is still tons of value in always telling AI to plan first, unless it's a small quick task. This post just takes it a few steps further and formalizes it.
I find Cursor works much more reliably using plan mode, reviewing/revising output in markdown, then pressing build. Which isn't a ton of overhead but often leads to lots of context switching as it definitely adds more time.
I partly agree with you. But once you have a codebase large enough, the changes become longer to even type in, once figured out.
I find the best way to use agents (and I don't use claude) is to hash it out like I'm about to write these changes and I make my own mental notes, and get the agent to execute on it.
Agents don't get tired, they don't start fat fingering stuff at 4pm, the quality doesn't suffer. And they can be parallelised.
Finally, this allows me to stay at a higher level and not get bogged down of "right oh did we do this simple thing again?" which wipes some of the context in my mind and gets tiring through the day.
Always, 100% review every line of code written by an agent though. I do not condone committing code you don't 'own'.
I'll never agree with a job that forces developers to use 'AI', I sometimes like to write everything by hand. But having this tool available is also very powerful.
I want to be clear, I'm not against any use of AI. It's hugely useful to save a couple of minutes of "write this specific function to do this specific thing that I could write and know exactly what it would look like". That's a great use, and I use it all the time! It's better autocomplete. Anything beyond that is pushing it - at the moment! We'll see, but spending all day writing specs and double-checking AI output is not more productive than just writing correct code yourself the first time, even if you're AI-autocompleting some of it.
For the last few days I've been working on a personal project that's been on ice for at least 6 years. Back when I first thought of the project and started implementing it, it took maybe a couple weeks to eke out some minimally working code.
This new version that I'm doing (from scratch with ChatGPT web) has a far more ambitious scope and is already at the "usable" point. Now I'm primarily solidifying things and increasing test coverage. And I've tested the key parts with IRL scenarios to validate that it's not just passing tests; the thing actually fulfills its intended function so far. Given the increased scope, I'm guessing it'd take me a few months to get to this point on my own, instead of under a week, and the quality wouldn't be where it is. Not saying I haven't had to wrangle with ChatGPT on a few bugs, but after a decent initial planning phase, my prompts now are primarily "Do it"s and "Continue"s. Would've likely already finished it if I wasn't copying things back and forth between browser and editor, and being forced to pause when I hit the message limit.
This is a great come-back story. I have had a similar experience with a photoshop demake of mine.
I recommend to try out Opencode with this approach, you might find it less tiring than ChatGPT web (yes it works with your ChatGPT Plus sub).
I think it comes down to "it depends". I work in a NIS2 regulated field and we're quite callenged by the fact that it means we can't give AI's any sort of real access because of the security risk. To be complaint we'd have to have the AI agent ask permission for every single thing it does, before it does it, and foureye review it. Which is obviously never going to happen. We can discuss how bad the NIS2 foureye requirement works in the real world another time, but considering how easy it is to break AI security, it might not be something we can actually ever use. This makes sense on some of the stuff we work on, since it could bring an entire powerplant down. On the flip-side AI risks would be of little concern on a lot of our internal tools, which are basically non-regulated and unimportant enough that they can be down for a while without costing the business anything beyond annoyances.
This is where our challenges are. We've build our own chatbot where you can "build" your own agent within the librechat framework and add a "skill" to it. I say "skill" because it's older than claude skills but does exactly the same. I don't completely buy the authors:
> “deeply”, “in great details”, “intricacies”, “go through everything”
bit, but you can obviously save a lot of time by writing a piece of english which tells it what sort of environment you work in. It'll know that when I write Python I use UV, Ruff and Pyrefly and so on as an example. I personally also have a "skill" setting that tells the AI not to compliment me because I find that ridicilously annoying, and that certainly works. So who knows? Anyway, employees are going to want more. I've been doing some PoC's running open source models in isolation on a raspberry pi (we had spares because we use them in IoT projects) but it's hard to setup an isolation policy which can't be circumvented.
We'll have to figure it out though. For powerplant critical projects we don't want to use AI. But for the web tool that allows a couple of employees to upload three excel files from an external accountant and then generate some sort of report on them? Who cares who writes it or even what sort of quality it's written with? The lifecycle of that tool will probably be something that never changes until the external account does and then the tool dies. Not that it would have necessarily been written in worse quality without AI... I mean... Have you seen some of the stuff we've written in the past 40 years?
There is a miscommunication happening, this entire time we all had surprisingly different ideas about what quality of work is acceptable which seems to account for differences of opinion on this stuff.
Surely Addy Osmani can code. Even he suggests plan first.
https://news.ycombinator.com/item?id=46489061
> planning and checking and prompting and orchestrating is far more work than just writing the code yourself.
This! Once I'm familiar with the codebase (which I strive to do very quickly), for most tickets, I usually have a plan by the time I've read the description. I can have a couple of implementation questions, but I knew where the info is located in the codebase. For things, I only have a vague idea, the whiteboard is where I go.
The nice thing with such a mental plan, you can start with a rougher version (like a drawing sketch). Like if I'm starting a new UI screen, I can put a placeholder text like "Hello, world", then work on navigation. Once that done, I can start to pull data, then I add mapping functions to have a view model,...
Each step is a verifiable milestone. Describing them is more mentally taxing than just writing the code (which is a flow state for me). Why? Because English is not fit to describe how computer works (try describe a finite state machine like navigation flow in natural languages). My mental mental model is already aligned to code, writing the solution in natural language is asking me to be ambiguous and unclear on purpose.
Shameless plug: https://beadhub.ai allows you to do exactly that, but with several agents in parallel. One of them is in the role of planner, which takes care of the source-of-truth document and the long term view. They all stay in sync with real-time chat and mail.
It's OSS.
Real-time work is happening at https://app.beadhub.ai/juanre/beadhub (beadhub is a public project at https://beadhub.ai so it is visible).
Particularly interesting (I think) is how the agents chat with each other, which you can see at https://app.beadhub.ai/juanre/beadhub/chat
The biggest roadblock to using agents to maximum effectiveness like this is the chat interface. It's convenience as detriment and convenience as distraction. I've found myself repeatedly giving into that convenience only to realize that I have wasted an hour and need to start over because the agent is just obliviously circling the solution that I thought was fully obvious from the context I gave it. Clearly these tools are exceptional at transforming inputs into outputs and, counterintuitively, not as exceptional when the inputs are constantly interleaved with the outputs like they are in chat mode.
What I've read is that even with all the meticulous planning, the author still needed to intervene. Not at the end but at the middle, unless it will continue building out something wrong and its even harder to fix once it's done. It'll cost even more tokens. It's a net negative.
You might say a junior might do the same thing, but I'm not worried about it, at least the junior learned something while doing that. They could do it better next time. They know the code and change it from the middle where it broke. It's a net positive.
Unfortunately, you could argue that the model provider has also learned something, i.e. the interaction can be used as additional training data to train subsequent models.
this comment is the first truly humane one ive read regarding this whole AI fiasco
Quoting the article:
> One trick I use constantly: for well-contained features where I’ve seen a good implementation in an open source repo, I’ll share that code as a reference alongside the plan request. If I want to add sortable IDs, I paste the ID generation code from a project that does it well and say “this is how they do sortable IDs, write a plan.md explaining how we can adopt a similar approach.” Claude works dramatically better when it has a concrete reference implementation to work from rather than designing from scratch.
Licensing apparently means nothing.
Ripped off in the training data, ripped off in the prompt.
Concepts are not copyrightable.
Well, that's already done by Amazon's Kiro [0], Google's Antigravity [1], GitHub's Spec Kit [2], and OpenSpec [3]!
[0]: https://kiro.dev/
[1]: https://antigravity.google/
[2]: https://github.github.com/spec-kit/
[3]: https://openspec.dev/
This is quite close to what I've arrived at, but with two modifications
1) anything larger I work on in layers of docs. Architecture and requirements -> design -> implementation plan -> code. Partly it helps me think and nail the larger things first, and partly helps claude. Iterate on each level until I'm satisfied.
2) when doing reviews of each doc I sometimes restart the session and clear context, it often finds new issues and things to clear up before starting the next phase.
Has anyone found a efficient way to avoid repeating the initial codebase assessment when working with large projects?
There are several projects on GitHub that attempt to tackle context and memory limitations, but I haven’t found one that consistently works well in practice.
My current workaround is to maintain a set of Markdown files, each covering a specific subsystem or area of the application. Depending on the task, I provide only the relevant documents to Claude Code to limit the context scope. It works reasonably well, but it still feels like a manual and fragile solution. I’m interested in more robust strategies for persistent project context or structured codebase understanding.
Whenever I build a new feature with it I end up with several plan files leftover. I ask CC to combine them all, update with what we actually ended up building and name it something sensible, then whenever I want to work on that area again it's a useful reference (including the architecture, decisions and tradeoffs, relevant files etc).
Yes this is what agent "skills" are. Just guides on any topic. The key is that you have the agent write and maintain them.
For my longer spec files, I grep the subheaders/headers (with line numbers) and show this compact representation to the LLM's context window. I also have a file that describes what each spec files is and where it's located, and I force the LLM to read that and pull the subsections it needs. I also have one entrypoint requirements file (20k tokens) that I force it to read in full before it does anything else, every line I wrote myself. But none of this is a silver bullet.
That sounds like the recommended approach. However, there's one more thing I often do: whenever Claude Code and I complete a task that didn't go well at first, I ask CC what it learned, and then I tell it to write down what it learned for the future. It's hard to believe how much better CC has become since I started doing that. I ask it to write dozens of unit tests and it just does. Nearly perfectly. It's insane.
I'm interested in this as well.
Skills almost seem like a solution, but they still need an out-of-band process to keep them updated as the codebase evolves. For now, a structured workflow that includes aggressive updates at the end of the loop is what I use.
In Claude Web you can use projects to put files relevant for context there.
I go a bit further than this and have had great success with 3 doc types and 2 skills:
- Specs: these are generally static, but updatable as the project evolves. And they're broken out to an index file that gives a project overview, a high-level arch file, and files for all the main modules. Roughly ~1k lines of spec for 10k lines of code, and try to limit any particular spec file to 300 lines. I'm intimately familiar with every single line in these.
- Plans: these are the output of a planning session with an LLM. They point to the associated specs. These tend to be 100-300 lines and 3 to 5 phases.
- Working memory files: I use both a status.md (3-5 items per phase roughly 30 lines overall), which points to a latest plan, and a project_status (100-200 lines), which tracks the current state of the project and is instructed to compact past efforts to keep it lean)
- A planner skill I use w/ Gemini Pro to generate new plans. It essentially explains the specs/plans dichotomy, the role of the status files, and to review everything in the pertinent areas of code and give me a handful of high-level next set of features to address based on shortfalls in the specs or things noted in the project_status file. Based on what it presents, I select a feature or improvement to generate. Then it proceeds to generate a plan, updates a clean status.md that points to the plan, and adjusts project_status based on the state of the prior completed plan.
- An implementer skill in Codex that goes to town on a plan file. It's fairly simple, it just looks at status.md, which points to the plan, and of course the plan points to the relevant specs so it loads up context pretty efficiently.
I've tried the two main spec generation libraries, which were way overblown, and then I gave superpowers a shot... which was fine, but still too much. The above is all homegrown, and I've had much better success because it keeps the context lean and focused.
And I'm only on the $20 plans for Codex/Gemini vs. spending $100/month on CC for half year prior and move quicker w/ no stall outs due to token consumption, which was regularly happening w/ CC by the 5th day. Codex rarely dips below 70% available context when it puts up a PR after an execution run. Roughly 4/5 PRs are without issue, which is flipped against what I experienced with CC and only using planning mode.
This is pretty much my approach. I started with some spec files for a project I'm working on right now, based on some academic papers I've written. I ended up going back and forth with Claude, building plans, pushing info back into the specs, expanding that out and I ended up with multiple spec/architecture/module documents. I got to the point where I ended up building my own system (using claude) to capture and generate artifacts, in more of a systems engineering style (e.g. following IEEE standards for conops, requirement documents, software definitions, test plans...). I don't use that for session-level planning; Claude's tools work fine for that. (I like superpowers, so far. It hasn't seemed too much)
I have found it to work very well with Claude by giving it context and guardrails. Basically I just tell it "follow the guidance docs" and it does. Couple that with intense testing and self-feedback mechanisms and you can easily keep Claude on track.
I have had the same experience with Codex and Claude as you in terms of token usage. But I haven't been happy with my Codex usage; Claude just feels like it's doing more of what I want in the way I want.
Looks good. Question - is it always better to use a monorepo in this new AI world? Vs breaking your app into separate repos? At my company we have like 6 repos all separate nextjs apps for the same user base. Trying to consolidate to one as it should make life easier overall.
It really depends but there’s nothing stopping you from just creating a separate folder with the cloned repositories (or worktrees) that you need and having a root CLAUDE.md file that explains the directory structure and referencing the individual repo CLAUDE.md files.
Just put all the repos in all in one directory yourself. In my experience that works pretty well.
AI is happy to work with any directory you tell it to. Agent files can be applied anywhere.
The multi-pass approach works outside of code too. I run a fairly complex automation pipeline (prompt -> script -> images -> audio -> video assembly) and the single biggest quality improvement was splitting generation into discrete planning and execution phases. One-shotting a 10-step pipeline means errors compound. Having the LLM first produce a structured plan, then executing each step against that plan with validation gates between them, cut my failure rate from maybe 40% to under 10%. The planning doc also becomes a reusable artifact you can iterate on without re-running everything.
> After Claude writes the plan, I open it in my editor and add inline notes directly into the document. These notes correct assumptions, reject approaches, add constraints, or provide domain knowledge that Claude doesn’t have.
This is the part that seems most novel compared to what I've heard suggested before. And I have to admit I'm a bit skeptical. Would it not be better to modify what Claude has written directly, to make it correct, rather than adding the corrections as separate notes (and expecting future Claude to parse out which parts were past Claude and which parts were the operator, and handle the feedback graciously)?
At least, it seems like the intent is to do all of this in the same session, such that Claude has the context of the entire back-and-forth updating the plan. But that seems a bit unpleasant; I would think the file is there specifically to preserve context between sessions.
One reason why I don't do this: even I won't be immune to mistakes. When I fix it with new values or paths, for example, and the one I provided is wrong, it can worsen the future work.
Personally, I like to order claude one more time to update the plan file after I have given annotation, and review it again after. This will ensure (from my understanding) that claude won't treat my annotation as different instructions, thus risking the work being conflicted.
Since everyone is showing their flow, here's mine:
* create a feature-name.md file in a gitignored folder
* start the file by giving the business context
* describe a high-level implementation and user flows
* describe database structure changes (I find it important not to leave it for interpretation)
* ask Claude to inspect the feature and review if for coherence, while answering its questions I ask to augment feature-name.md file with the answers
* enter Claude's plan mode and provide that feature-name.md file
* at this point it's detailed enough that rarely any corrections from me are needed
Planning is important because you get the LLM to explain the problem and solution in its language and structure, not yours.
This shortcuts a range of problem cases where the LLM fights between the users strict and potentially conflicting requirements, and its own learning.
In the early days we used to get LLM to write the prompts for us to get round this problem, now we have planning built in.
I try these staging-document patterns, but suspect they have 2 fundamental flaws that stem mostly from our own biases.
First, Claude evolves. The original post work pattern evolved over 9 months, before claude's recent step changes. It's likely claude's present plan mode is better than this workaround, but if you stick to the workaround, you'd never know.
Second, the staging docs that represent some context - whether a library skills or current session design and implementation plans - are not the model Claude works with. At best they are shaping it, but I've found it does ignore and forget even what's written (even when I shout with emphasis), and the overall session influences the code. (Most often this happens when a peripheral adjustment ends up populating half the context.)
Indeed the biggest benefit from the OP might be to squeeze within 1 session, omitting peripheral features and investigations at the plan stage. So the mechanism of action might be the combination of getting our own plan clear and avoiding confusing excursions. (A test for that would be to redo the session with the final plan and implementation, to see if the iteration process itself is shaping the model.)
Our bias is to believe that we're getting better at managing this thing, and that we can control and direct it. It's uncomfortable to realize you can only really influence it - much like giving direction to a junior, but they can still go off track. And even if you found a pattern that works, it might work for reasons you're not understanding -- and thus fail you eventually. So, yes, try some patterns, but always hang on to the newbie senses of wonder and terror that make you curious, alert, and experimental.
This is the flow I've found myself working towards. Essentially maintaining more and more layered documentation for the LLM produces better and more consistent results. What is great here is the emphasis on the use of such documents in the planning phase. I'm feeling much more motivated to write solid documentation recently, because I know someone (the LLM) is actually going to read it! I've noticed my efforts and skill acquisition have moved sharply from app developer towards DevOps and architecture / management, but I think I'll always be grateful for the application engineering experience that I think the next wave of devs might miss out on.
I've also noted such a huge gulf between some developers describing 'prompting things into existence' and the approach described in this article. Both types seem to report success, though my experience is that the latter seems more realistic, and much more likely to produce robust code that's likely to be maintainable for long term or project critical goals.
The separation of planning and execution resonates strongly. I've been using a similar pattern when building with AI APIs — write the spec/plan in natural language first, then let the model execute against it.
One addition that's worked well for me: keeping a persistent context file that the model reads at the start of each session. Instead of re-explaining the project every time, you maintain a living document of decisions, constraints, and current state. Turns each session into a continuation rather than a cold start.
The biggest productivity gain isn't in the code generation itself — it's in reducing the re-orientation overhead between sessions.
This is the way.
The practice is:
- simple
- effective
- retains control and quality
Certainly the “unsupervised agent” workflows are getting a lot of attention right now, but they require a specific set of circumstances to be effective:
- clear validation loop (eg. Compile the kernel, here is gcc that does so correctly)
- ai enabled tooling (mcp / cli tool that will lint, test and provide feedback immediately)
- oversight to prevent sgents going off the rails (open area of research)
- an unlimited token budget
That means that most people can't use unsupervised agents.
Not that they dont work; Most people have simply not got an environment and task that is appropriate.
By comparison, anyone with cursor or claude can immediately start using this approach, or their own variant on it.
It does not require fancy tooling.
It does not require an arcane agent framework.
It works generally well across models.
This is one of those few genunie pieces of good practical advice for people getting into AI coding.
Simple. Obviously works once you start using it. No external dependencies. BYO tools to help with it, no “buy my AI startup xxx to help”. No “star my github so I can a job at $AI corp too”.
Great stuff.
Honesty this is just language models in general at the moment, and not just coding.
It’s the same reason adding a thinking step works.
You want to write a paper, you have it form a thesis and structure first. (In this one you might be better off asking for 20 and seeing if any of them are any good.) You want to research something, first you add gathering and filtering steps before synthesis.
Adding smarter words or telling it to be deeper does work by slightly repositioning where your query ends up in space.
Asking for the final product first right off the bat leads to repetitive verbose word salad. It just starts to loop back in on itself. Which is why temperature was a thing in the first place, and leads me to believe they’ve turned the temp down a bit to try and be more accurate. Add some randomness and variability to your prompts to compensate.
Absolutely. And you can also always let the agent look back at the plan to check if it is still on track and aligned.
One step I added, that works great for me, is letting it write (api-level) tests after planning and before implementation. Then I’ll do a deep review and annotation of these tests and tweak them until everything is just right.
Huge +1. This loop consistently delivers great results for my vibe coding.
The “easy” path of “short prompt declaring what I want” works OK for simple tasks but consistently breaks down for medium to high complexity tasks.
Can you help me understand the difference between "short prompt for what I want (next)" vs medium to high complexity tasks?
What i mean is, in practice, how does one even get to a a high complexity task? What does that look like? Because isn't it more common that one sees only so far ahead?
It's more or less what comes out of the box with plan mode, plus a few extra bits?
This is similar to what I do. I instruct an Architect mode with a set of rules related to phased implementation and detailed code artifacts output to a report.md file. After a couple of rounds of review and usually some responses that either tie together behaviors across context, critique poor choices or correct assumptions, there is a piece of work defined for a coder LLM to perform. With the new Opus 4.6 I then select specialist agents to review the report.md, prompted with detailed insight into particular areas of the software. The feedback from these specialist agent reviews is often very good and sometimes catches things I had missed. Once all of this is done, I let the agent make the changes and move onto doing something else. I typically rename and commit the report.md files which can be useful as an alternative to git diff / commit messages etc.
Good article, but I would rephrase the core principle slightly:
Never let Claude write code until you’ve reviewed, *fully understood* and approved a written plan.
In my experience, the beginning of chaos is the point at which you trust that Claude has understood everything correctly and claims to present the very best solution. At that point, you leave the driver's seat.
This is what I do with the obra/superpowers[0] set of skills.
1. Use brainstorming to come up with the plan using the Socratic method
2. Write a high level design plan to file
3. I review the design plan
4. Write an implementation plan to file. We've already discussed this in detail, so usually it just needs skimming.
5. Use the worktree skill with subagent driven development skill
6. Agent does the work using subagents that for each task:
7. When all tasks complete: create a PR for me to review8. Go back to the agent with any comments
9. If finished, delete the plan files and merge the PR
[0]: https://github.com/obra/superpowers
If you’ve ever desired the ability for annotating the plan more visually, try fitting Plannotator in this workflow. There is a slash command for use when you use custom workflows outside of normal plan mode.
https://github.com/backnotprop/plannotator
I'll give this a try. Thanks for the suggestion.
The crowd around this pot shows how superficial is knowledge about claude code. It gets releases each day and most of this is already built in the vanilla version. Not to mention subagent working in work trees, memory.md, plan on which you can comment directly from the interface, subagents launched in research phase, but also some basic mcp's like LSP/IDE integration, and context7 to not to be stuck in the knowledge cutoff/past.
When you go to YouTube and search for stuff like "7 levels of claude code" this post would be maybe 3-4.
Oh, one more thing - quality is not consistent, so be ready for 2-3 rounds of "are you happy with the code you wrote" and defining audit skills crafted for your application domain - like for example RODO/Compliance audit etc.
I'm using the in-built features as well, but I like the flow that I have with superpowers. You've made a lot of assumptions with your comment that are just not true (at least for me).
I find that brainstorming + (executing plans OR subagent driven development) is way more reliable than the built-in tooling.
I had to stop reading about half way, it's written in that breathless linkedin/ai generated style.
I do something very similar, also with Claude and Codex, because the workflow is controlled by me, not by the tool. But instead of plan.md I use a ticket system basically like ticket_<number>_<slug>.md where I let the agent create the ticket from a chat, correct and annotate it afterwards and send it back, sometimes to a new agent instance. This workflow helps me keeping track of what has been done over time in the projects I work on. Also this approach does not need any „real“ ticket system tooling/mcp/skill/whatever since it works purely on text files.
+1 to creating tickets by simply asking the agent to. It's worked great and larger tasks can be broken down into smaller subtasks that could reasonably be completed in a single context window, so you rarely every have to deal with compaction. Especially in the last few months since Claude's gotten good at dispatching agents to handle tasks if you ask it to, I can plan large changes that span multilpe tickets and tell claude to dispatch agents as needed to handle them (which it will do in parallel if they mostly touch different files), keeping the main chat relatively clean for orchestration and validation work.
semantic plan name is important
I’ve been using this same pattern, except not the research phase. Definetly will try to add it to my process aswell.
Sometimes when doing big task I ask claude to implement each phase seprately and review the code after each step.
Regarding inline notes, I use a specific format in the `/plan` command, by using th `ME:` prefix.
https://github.com/srid/AI/blob/master/commands/plan.md#2-pl...
It works very similar to Antigravity's plan document comment-refine cycle.
https://antigravity.google/docs/implementation-plan
There are frameworks like https://github.com/bmad-code-org/BMAD-METHOD and https://github.github.com/spec-kit/ that are working on encoding a similar kind of approach and process.
I've been working off and on on a vibe coded FP language and transpiler - mostly just to get more experience with Claude Code and see how it handles complex real world projects. I've settled on a very similar flow, though I use three documents: plan, context, task list. Multiple rounds of iteration when planning a feature. After completion, have a clean session do an audit to confirm that everything was implemented per the design. Then I have both Claude and CodeRabbit do code review passes before I finally do manual review. VERY heavy emphasis on tests, the project currently has 2x more test code than application code. So far it works surprisingly well. Example planning docs below -
https://github.com/mbcrawfo/vibefun/tree/main/.claude/archiv...
The annotation cycle is the key insight for me. Treating the plan as a living doc you iterate on before touching any code makes a huge difference in output quality.
Experimentally, i've been using mfbt.ai [https://mfbt.ai] for roughly the same thing in a team context. it lets you collaboratively nail down the spec with AI before handing off to a coding agent via MCP.
Avoids the "everyone has a slightly different plan.md on their machine" problem. Still early days but it's been a nice fit for this kind of workflow.
I agree, and this is why I tend to use gptel in emacs for planning - the document is the conversation context, and can be edited and annotated as you like.
In my own tests I have found opus to be very good at writing plans, terrible at executing them. It typically ignores half of the constraints. https://x.com/xundecidability/status/2019794391338987906?s=2... https://x.com/xundecidability/status/2024210197959627048?s=2...
1. Don't implement too much at at time
2. Have the agent review if it followed the plan and relevant skills accurately.
the first link was from a simple request with fewer than 1000 tokens total in the context window, just a short shell script.
here is another one which had about 200 tokens and opus decided to change the model name i requested.
https://x.com/xundecidability/status/2005647216741105962?s=2...
opus is bad at instruction following now.
I've been teaching AI coding tool workshops for the past year and this planning-first approach is by far the most reliable pattern I've seen across skill levels.
The key insight that most people miss: this isn't a new workflow invented for AI - it's how good senior engineers already work. You read the code deeply, write a design doc, get buy-in, then implement. The AI just makes the implementation phase dramatically faster.
What I've found interesting is that the people who struggle most with AI coding tools are often junior devs who never developed the habit of planning before coding. They jump straight to "build me X" and get frustrated when the output is a mess. Meanwhile, engineers with 10+ years of experience who are used to writing design docs and reviewing code pick it up almost instantly - because the hard part was always the planning, not the typing.
One addition I'd make to this workflow: version your research.md and plan.md files in git alongside your code. They become incredibly valuable documentation for future maintainers (including future-you) trying to understand why certain architectural decisions were made.
Every "how I use Claude Code" post will get into the HN frontpage.
Which maybe has to do with people wanting to show how they use Claude Code in the comments!
I have to give this a try. My current model for backend is the same as how author does frontend iteration. My friend does the research-plan-edit-implement loop, and there is no real difference between the quality of what I do and what he does. But I do like this just for how it serves as documentation of the thought process across AI/human, and can be added to version control. Instead of humans reviewing PRs, perhaps humans can review the research/plan document.
On the PR review front, I give Claude the ticket number and the branch (or PR) and ask it to review for correctness, bugs and design consistency. The prompt is always roughly the same for every PR. It does a very good job there too.
Modelwise, Opus 4.6 is scary good!
Lol I wrote about this and been using plan+execute workflow for 8 months.
Sadly my post didn't much attention at the time.
https://thegroundtruth.media/p/my-claude-code-workflow-and-p...
The author is quite far on their journey but would benefit from writing simple scripts to enforce invariants in their codebase. Invariant broken? Script exits with a non-zero exit code and some output that tells the agent how to address the problem. Scripts are deterministic, run in milliseconds, and use zero tokens. Put them in husky or pre-commit, install the git hooks, and your agent won’t be able to commit without all your scripts succeeding.
And “Don’t change this function signature” should be enforced not by anticipating that your coding agent “might change this function signature so we better warn it not to” but rather via an end to end test that fails if the function signature is changed (because the other code that needs it not to change now has an error). That takes the author out of the loop and they can not watch for the change in order to issue said correction, and instead sip coffee while the agent observes that it caused a test failure then corrects it without intervention, probably by rolling back the function signature change and changing something else.
https://github.blog/ai-and-ml/generative-ai/spec-driven-deve...
Haha this is surprisingly and exactly how I use claude as well. Quite fascinating that we independently discovered the same workflow.
I maintain two directories: "docs/proposals" (for the research md files) and "docs/plans" (for the planning md files). For complex research files, I typically break them down into multiple planning md files so claude can implement one at a time.
A small difference in my workflow is that I use subagents during implementation to avoid context from filling up quickly.
Same, I formalized a similar workflow for my team (oriented around feature requirement docs), I am thinking about fully productizing it and am looking to for feedback - https://acai.sh
Even if the product doesn’t resonate I think I’ve stumbled on some ideas you might find useful^
I do think spec-driven development is where this all goes. Still making up my mind though.
Spec-driven looks very much like what the author describes. He may have some tweaks of his own but they could just as well be coded into the artifacts that something like OpenSpec produces.
This is basically long-lived specs that are used as tests to check that the product still adheres to the original idea that you wanted to implement, right?
This inspired me to finally write good old playwright tests for my website :).
I don't deny that AI has use cases, but boy - the workflow described is boring:
"Most developers type a prompt, sometimes use plan mode, fix the errors, repeat. "
Does anyone think this is as epic as, say, watch the Unix archives https://www.youtube.com/watch?v=tc4ROCJYbm0 where Brian demos how pipes work; or Dennis working on C and UNIX? Or even before those, the older machines?
I am not at all saying that AI tools are all useless, but there is no real epicness. It is just autogenerated AI slop and blob. I don't really call this engineering (although I also do agree, that it is engineering still; I just don't like using the same word here).
> never let Claude write code until you’ve reviewed and approved a written plan.
So the junior-dev analogy is quite apt here.
I tried to read the rest of the article, but I just got angrier. I never had that feeling watching oldschool legends, though perhaps some of their work may be boring, but this AI-generated code ... that's just some mythical random-guessing work. And none of that is "intelligent", even if it may appear to work, may work to some extent too. This is a simulation of intelligence. If it works very well, why would any software engineer still be required? Supervising would only be necessary if AI produces slop.
> “remove this section entirely, we don’t need caching here” — rejecting a proposed approach
I wonder why you don't remove it yourself. Aren't you already editing the plan?
Interesting! I feel like I'm learning to code all over again! I've only been using Claude for a little more than a month and until now I've been figuring things out on my own. Building my methodology from scratch. This is much more advanced than what I'm doing. I've been going straight to implementation, but doing one very small and limited feature at a time, describing implementation details (data structures like this, use that API here, import this library etc) verifying it manually, and having Claude fix things I don't like. I had just started getting annoyed that it would make the same (or very similar) mistake over and over again and I would have to fix it every time. This seems like it'll solve that problem I had only just identified! Neat!
this is exactly how I work with cursor
except that I put notes to plan document in a single message like:
otherwise, I'm not sure how to guarantee that ai won't confuse my notes with its own plan.one new thing for me is to review the todo list, I was always relying on auto generated todo list
> Most developers type a prompt, sometimes use plan mode, fix the errors, repeat.
> ...
> never let Claude write code until you’ve reviewed and approved a written plan
I certainly always work towards an approved plan before I let it lost on changing the code. I just assumed most people did, honestly. Admittedly, sometimes there's "phases" to the implementation (because some parts can be figured out later and it's more important to get the key parts up and running first), but each phase gets a full, reviewed plan before I tell it to go.
In fact, I just finished writing a command and instruction to tell claude that, when it presents a plan for implementation, offer me another option; to write out the current (important parts of the) context and the full plan to individual (ticket specific) md files. That way, if something goes wrong with the implementation I can tell it to read those files and "start from where they left off" in the planning.
The author seems to think theyve invented a special workflow...
We all tend to regress to average (same thoughts/workflows)...
Have had many users already doing the exact same workflow with: https://github.com/backnotprop/plannotator
4 times in one thread, please stop spamming this link.
I recently discovered GitHub speckit which separates planning/execution in stages: specify, plan, tasks, implement. Finding it aligns with the OP with the level of “focus” and “attention” this gets out of Claude Code.
Speckit is worth trying as it automates what is being described here, and with Opus 4.6 it's been a kind of BC/AD moment for me.
Try OpenSpec and it'll do all this for you. SpecKit works too. I don't think there's a need to reinvent the wheel on this one, as this is spec-driven development.
I’ve been using Claude through opencode, and I figured this was just how it does it. I figured everyone else did it this way as well. I guess not!
Since the rise of AI systems I really wonder how people wrote code before. This is exactly how I planned out implementation and executed the plan. Might have been some paper notes, a ticket or a white board, buuuuut ... I don't know.
Cool, the idea of leaving comments directly in the plan never even occurred to me, even though it really is the obvious thing to do.
Do you markup and then save your comments in any way, and have you tried keeping them so you can review the rules and requirements later?
I tried Opus 4.6 recently and it’s really good. I had ditched Claude a long time ago for Grok + Gemini + OpenCode with Chinese models. I used Grok/Gemini for planning and core files, and OpenCode for setup, running, deploying, and editing.
However, Opus made me rethink my entire workflow. Now, I do it like this:
* PRD (Product Requirements Document)
* main.py + requirements.txt + readme.md (I ask for minimal, functional, modular code that fits the main.py)
* Ask for a step-by-step ordered plan
* Ask to focus on one step at a time
The super powerful thing is that I don’t get stuck on missing accounts, keys, etc. Everything is ordered and runs smoothly. I go rapidly from idea to working product, and it’s incredibly easy to iterate if I figure out new features are required while testing. I also have GLM via OpenCode, but I mainly use it for "dumb" tasks.
Interestingly, for reasoning capabilities regarding standard logic inside the code, I found Gemini 3 Flash to be very good and relatively cheap. I don't use Claude Code for the actual coding because forcing everything via chat into a main.py encourages minimal code that's easy to skim—it gives me a clearer representation of the feature space
Interesting approach. The separation of planning and execution is crucial, but I think there's a missing layer most people overlook: permission boundaries between the two phases.
Right now when Claude Code (or any agent) executes a plan, it typically has the same broad permissions for every step. But ideally, each execution step should only have access to the specific tools and files it needs — least privilege, applied to AI workflows.
I've been experimenting with declarative permission manifests for agent tasks. Instead of giving the agent blanket access, you define upfront what each skill can read, write, and execute. Makes the planning phase more constrained but the execution phase much safer.
Anyone else thinking about this from a security-first angle?
I’m a big fan of having the model create a GitHub issue directly (using the GH CLI) with the exact plan it generates, instead of creating a markdown file that will eventually get deleted. It gives me a permanent record and makes it easy to reference and close the issue once the PR is ready.
Google Anti-Gravity has this process built in. This is essentially a cycle a developer would follow: plan/analyse - document/discuss - break down tasks/implement. We’ve been using requirements and design documents as best practice since leaving our teenage bedroom lab for the professional world. I suppose this could be seen as our coding agents coming of age.
All sounds like a bespoke way of remaking https://github.com/Fission-AI/OpenSpec
My process is similar, but I recently added a new "critique the plan" feedback loop that is yielding good results. Steps:
1. Spec
2. Plan
3. Read the plan & tell it to fix its bad ideas.
4. (NB) Critique the plan (loop) & write a detailed report
5. Update the plan
6. Review and check the plan
7. Implement plan
Detailed here:
https://x.com/PetrusTheron/status/2016887552163119225
Same. In my experience, the first plan always benefits from being challenged once or twice by claude itself.
I came to the exact same pattern, with one extra heuristic at the end: spin up a new claude instance after the implementation is complete and ask it to find discrepancies between the plan and the implementation.
> I am not seeing the performance degradation everyone talks about after 50% context window.
I pretty much agree with that. I use long sessions and stopped trying to optimize the context size, the compaction happens but the plan keeps the details and it works for me.
The baffling part of the article is all the assertions about how this is unique, novel, not the typical way people are doing this etc.
There are whole products wrapped around this common workflow already (like Augment Intent).
Doesn’t Claude code do this by switching between edit mode and plan mode?
FWIW I have had significant improvements by clearing context then implementing the plan. Seems like it stops Claude getting hung up on something.
I agree with most of this, though I'm not sure it's radically different. I think most people who've been using CC in earnest for a while probably have a similar workflow? Prior to Claude 4 it was pretty much mandatory to define requirements and track implementation manually to manage context. It's still good, but since 4.5 release, it feels less important. CC basically works like this by default now, so unless you value the spec docs (still a good reference for Claude, but need to be maintained), you don't have to think too hard about it anymore.
The important thing is to have a conversation with Claude during the planning phase and don't just say "add this feature" and take what you get. Have a back and forth, ask questions about common patterns, best practices, performance implications, security requirements, project alignment, etc. This is a learning opportunity for you and Claude. When you think you're done, request a final review to analyze for gaps or areas of improvement. Claude will always find something, but starts to get into the weeds after a couple passes.
If you're greenfield and you have preferences about structure and style, you need to be explicit about that. Once the scaffolding is there, modern Claude will typically follow whatever examples it finds in the existing code base.
I'm not sure I agree with the "implement it all without stopping" approach and let auto-compact do its thing. I still see Claude get lazy when nearing compaction, though has gotten drastically better over the last year. Even so, I still think it's better to work in a tight loop on each stage of the implementation and preemptively compacting or restarting for the highest quality.
Not sure that the language is that important anymore either. Claude will explore existing codebase on its own at unknown resolution, but if you say "read the file" it works pretty well these days.
My suggestions to enhance this workflow:
- If you use a numbered phase/stage/task approach with checkboxes, it makes it easy to stop/resume as-needed, and discuss particular sections. Each phase should be working/testable software.
- Define a clear numbered list workflow in CLAUDE.md that loops on each task (run checks, fix issues, provide summary, etc).
- Use hooks to ensure the loop is followed.
- Update spec docs at the end of the cycle if you're keeping them. It's not uncommon for there to be some divergence during implementation and testing.
Sounds a bit like what Claude Plan Mode or Amazon's Kiro were built for. I agree it's a useful flow, but you can also overdo it.
I don't really get what is different about this from how almost everyone else uses Claude Code? This is an incredibly common, if not the most common way of using it (and many other tools).
Another approach is to spec functionality using comments and interfaces, then tell the LLM to first implement tests and finally make the tests pass. This way you also get regression safety and can inspect that it works as it should via the tests.
this sounds... really slow. for large changes for sure i'm investing time into planning. but such a rigid system can't possible be as good as a flexible approach with variable amounts of planning based on complexity
This is a similar workflow to speckit, kiro, gsd, etc.
Insights are nice for new users but I’m not seeing anything too different from how anyone experienced with Claude Code would use plan mode. You can reject plans with feedback directly in the CLI.
It is really fun to watch how a baby makes its first steps and also how experienced professionals rediscover what standards were telling us for 80+ years.
How are the annotations put into the markdown? Claude needs to be able to identify them as annotations and not parts of the plan.
It seems like the annotation of plan files is the key step.
Claude Code now creates persistent markdown plan files in ~/.claude/plans/ and you can open them with Ctrl-G to annotate them in your default editor.
So plan mode is not ephemeral any more.
Has Claude Code become slow, laggy, imprecise, giving wrong answers for other people here?
I use amazon kiro.
The AI first works with you to write requirements, then it produces a design, then a task list.
The helps the AI to make smaller chunks to work on, it will work on one task at a time.
I can let it run for an hour or more in this mode. Then there is lots of stuff to fix, but it is mostly correct.
Kiro also supports steering files, they are files that try to lock the AI in for common design decisions.
the price is that a lot of the context is used up with these files and kiro constantly pauses to reset the context.
This is exactly how I use it.
Funny how I came up with something loosely similar. Asking Codex to write a detailed plan in a markdown document, reviewing it, and asking it to implement it step by step. It works exquisitely well when it can build and test itself.
What works extremely well for me is this: Let Claude Code create the plan, then turn over the plan to Codex for review, and give the response back to Claude Code. Codex is exceptionally good at doing high level reviews and keeping an eye on the details. It will find very suble errors and omissins. And CC is very good at quickly converting the plan into code.
This back and forth between the two agents with me steering the conversation elevates Claude Code into next level.
Hub and spoke documentation in planning has been absolutely essential for the way my planning was before, and it's pretty cool seeing it work so well for planning mode to build scaffolds and routing.
Why don't you make Claude give feedback and iterate by itself?
this is literally reinventing claude's planning mode, but with more steps. I think Boris doesn't realize that planning mode is actually stored in a file.
https://x.com/boristane/status/2021628652136673282
I have tried using this and other workflows for a long time and had never been able to get them to work (see chat history for details).
This has changed in the last week, for 3 reasons:
1. Claude opus. It’s the first model where I haven’t had to spend more time correcting things than it would’ve taken me to just do it myself. The problem is that opus chews through tokens, which led to..
2. I upgraded my Claude plan. Previously on the regular plan I’d get about 20 mins of time before running out of tokens for the session and then needing to wait a few hours to use again. It was fine for little scripts or toy apps but not feasible for the regular dev work I do. So I upgraded to 5x. This now got me 1-2 hours per session before tokens expired. Which was better but still a frustration. Wincing at the price, I upgraded again to the 20x plan and this was the next game changer. I had plenty of spare tokens per session and at that price it felt like they were being wasted - so I ramped up my usage. Following a similar process as OP but with a plans directory with subdirectories for backlog, active and complete plans, and skills with strict rules for planning, implementing and completing plans, I now have 5-6 projects on the go. While I’m planning a feature on one the others are implementing. The strict plans and controls keep them on track and I have follow up skills for auditing quality and performance. I still haven’t hit token limits for a session but I’ve almost hit my token limit for the week so I feel like I’m getting my money’s worth. In that sense spending more has forced me to figure out how to use more.
3. The final piece of the puzzle is using opencode over claude code. I’m not sure why but I just don’t gel with Claude code. Maybe it’s all the sautéing and flibertygibbering, maybe it’s all the permission asking, maybe it’s that it doesn’t show what it’s doing as much as opencode. Whatever it is it just doesn’t work well for me. Opencode on the other hand is great. It’s shows what it’s doing and how it’s thinking which makes it easy for me to spot when it’s going off track and correct early.
Having a detailed plan, and correcting and iterating on the plan is essential. Making clause follow the plan is also essential - but there’s a line. Too fine grained and it’s not as creative at solving problems. Too loose/high level and it makes bad choices and goes in the wrong direction.
Is it actually making me more productive? I think it is but I’m only a week in. I’ve decided to give myself a month to see how it all works out.
I don’t intend to keep paying for the 20x plan unless I can see a path to using it to earn me at least as much back.
Just don’t use Claude Code. I can use the Codex CLI with just my $20 subscription and never come close to any usage limits
What if it's just slower so that your daily work fits within the paid tier they want?
It isn’t slower. I use my personal ChatGPT subscriptions with Codex for almost everything at work and use my $800/month company Claude allowance only for the tricky stuff that Codex can’t figure out. It’s never application code. It’s usually some combination of app code + Docker + AWS issue with my underlying infrastructure - created with whatever IAC that I’m using for a client - Terraform/CloudFormation or the CDK.
I burned through $10 on Claude in less than an hour. I only have $36 a day at $800 a month (800/22 working days)
> and use my $800/month company Claude allowance only for the tricky stuff that Codex can’t figure out.
It doesn’t seem controversial that the model that can solve more complex problems (that you admit the cheaper model can’t solve) costs more.
For the things I use it for, I’ve not found any other model to be worth it.
You’re assuming rational behavior from a company that doesn’t care about losing billions of dollar.
Have you tried Codex with OpenAi’s latest models?
Not in the last 2 months.
Current clause subscription is a sunk cost for the next month. Maybe I’ll try codex if Claude doesn’t lead anywhere.
I use both. As I’m working, I tell each of them to update a common document with the conversation. I don’t just tell Claude the what. I tell it the why and have it document it.
I can switch back and forth and use the MD file as shared context.
Curious: what are some cases where it'd make sense to not pay for the 20x plan (which is $200/month), and provide a whopping $800/month pay-per-token allowance instead?
Who knows? It’s part of an enterprise plan. I work for a consulting company. There are a number of fallbacks, the first fallback if we are working on an internal project is just to use our internal AWS account and use Claude code with the Anthropic hosted on Bedrock.
https://code.claude.com/docs/en/amazon-bedrock
The second fallback if it is for a customer project is to use their AWS account for development for them.
The rate my company charges for me - my level as an American based staff consultant (highest bill rate at the company) they are happy to let us use Claude Code using their AWS credentials. Besides, if we are using AWS Bedrock hosted Anthropic models, they know none of their secrets are going to Anthropic. They already have the required legal confidentiality/compliancd agreements with AWS.
It’s worrying to me that nobody really knows how LLMs work. We create prompts with or without certain words and hope it works. That’s my perspective anyway
It's actually no different from how real software is made. Requirements come from the business side, and through an odd game of telephone get down to developers.
The team that has developers closest to the customer usually makes the better product...or has the better product/market fit.
Then it's iteration.
It's the same as dealing with a human. You convey a spec for a problem and the language you use matters. You can convey the problem in (from your perspective) a clear way and you will get mixed results nonetheless. You will have to continue to refine the solution with them.
Genuinely: no one really knows how humans work either.
There are a few prompt frameworks that essentially codify these types of workflows by adding skills and prompts
https://github.com/obra/superpowers https://github.com/jlevy/tbd
Is it required to tell Claude to re-read the code folder again when you come back some day later or should we ask Claude to just pickup from research.md file thus saving some tokens?
The “inline comments on a plan” is one of the best features of Antigravity, and I’m surprised others haven’t started copycatting.
I feel like if I have to do all this, I might as well write the code myself.
Sorry but I didn't get the hype with this post, isnt it what most of the people doing? I want to see more posts on how you use the claude "smart" without feeding the whole codebase polluting the context window and also more best practices on cost efficient ways to use it, this workflow is clearly burning million tokens per session, for me is a No
That is just spec driven development without a spec, starting with the plan step instead.
I do something broadly similar. I ask for a design doc that contains an embedded todo list, broken down into phases. Looping on the design doc asking for suggestions seems to help. I'm up to about 40 design docs so far on my current project.
Claude appeared to just crash in my session: https://news.ycombinator.com/item?id=47107630
This is just Waterfall for LLMs. What happens when you explore the problem space and need to change up the plan?
That's exactly what Cursor's "plan" mode does? It even creates md files, which seems to be the main "thing" the author discovered. Along with some cargo cult science?
How is this noteworthy other than to spark a discussion on hn? I mean I get it, but a little more substance would be nice.
This is great. My workflow is also heading in that direction, so this is a great roadmap. I've already learned that just naively telling Claude what to do and letting it work, is a recipe for disaster and wasted time.
I'm not this structured yet, but I often start with having it analyse and explain a piece of code, so I can correct it before we move on. I also often switch to an LLM that's separate from my IDE because it tends to get confused by sprawling context.
AI only improves and changes. Embrace the scientific method and make sure your “here’s how to” are based in data.
my rlm-workflow skill has this encoded as a repeatable workflow.
give it a try: https://skills.sh/doubleuuser/rlm-workflow/rlm-workflow
That's great, actually, doesn't the logic apply to other services as well?
My workflow is a bit different.
* I ask the LLM for it's understanding of a topic or an existing feature in code. It's not really planning, it's more like understanding the model first
* Then based on its understanding, I can decide how great or small to scope something for the LLM
* An LLM showing good understand can deal with a big task fairly well.
* An LLM showing bad understanding still needs to be prompted to get it right
* What helps a lot is reference implementations. Either I have existing code that serves as the reference or I ask for a reference and I review.
A few folks do it at my work do it OPs way, but my arguments for not doing it this way
* Nobody is measuring the amount of slop within the plan. We only judge the implementation at the end
* it's still non deterministic - folks will have different experiences using OPs methods. If claude updates its model, it outdates OPs suggestions by either making it better or worse. We don't evaluate when things get better, we only focus on things not gone well.
* it's very token heavy - LLM providers insist that you use many tokens to get the task done. It's in their best interest to get you to do this. For me, LLMs should be powerful enough to understand context with minimal tokens because of the investment into model training.
Both ways gets the task done and it just comes down to my preference for now.
For me, I treat the LLM as model training + post processing + input tokens = output tokens. I don't think this is the best way to do non deterministic based software development. For me, we're still trying to shoehorn "old" deterministic programming into a non deterministic LLM.
Is this not just Ralph with extra steps and the risk of context rot?
How much time are you actually saving at this point?
Tip: LLMs are very good at following conventions (this is actually what is happening when it writes code). If you create a .md file with a list of entries of the following structure: # <identifier> <description block> <blank space> # <identifier> ... where an <identifier> is a stable and concise sequence of tokens that identifies some "thing" and seed it with 5 entries describing abstract stuff, the LLM will latch on and reference this. I call this a PCL (Project Concept List). I just tell it: > consume tmp/pcl-init.md pcl.md The pcl-init.md describes what PCL is and pcl.md is the actual list. I have pcl.md file for each independent component in the code (logging, http, auth, etc). This works very very well. The LLM seems to "know" what you're talking about. You can ask questions and give instructions like "add a PCL entry about this". It will ask if should add a PCL entry about xyz. If the description block tends to be high information-to-token ratio, it will follow that convention (which is a very good convention BTW).
However, there is a caveat. LLMs resist ambiguity about authority. So the "PCL" or whatever you want to call it, needs to be the ONE authoritative place for everything. If you have the same stuff in 3 different files, it won't work nearly as well.
Bonus Tip: I find long prompt input with example code fragments and thoughtful descriptions work best at getting an LLM to produce good output. But there will always be holes (resource leaks, vulnerabilities, concurrency flaws, etc). So then I update my original prompt input (keep it in a separate file PROMPT.txt as a scratch pad) to add context about those things maybe asking questions along the way to figure out how to fix the holes. Then I /rewind back to the prompt and re-enter the updated prompt. This feedback loop advances the conversation without expending tokens.
You described how AntiGravity works natively.
falling asleep here. when will the babysitting end
Use OpenSpec and simplify everything.
I don't know. I tried various methods. And this one kind of doesn't work quite a bit of times. The problem is plan naturally always skips some important details, or assumes some library function, but is taken as instruction in the next section. And claude can't handle ambiguity if the instruction is very detailed(e.g. if plan asks to use a certain library even if it is a bad fit claude won't know that decision is flexible). If the instruction is less detailed, I saw claude is willing to try multiple things and if it keeps failing doesn't fear in reverting almost everything.
In my experience, the best scenario is that instruction and plan should be human written, and be detailed.
Honestly, I found that the best way to use these CLIs is exactly how the CLI creators have intended.
We're just slowly reinventing agile for telling Ai agents what to do lol
Just skip to the Ai stand-ups
The plan document and todo are an artifact of context size limits. I use them too because it allows using /reset and then continuing.
Another pattern is:
1. First vibecode software to figure out what you want
2. Then throw it out and engineer it
Wow, I never bother with using phrases like “deeply study this codebase deeply.” I consistently get pretty fantastic results.
I have a different approach where I have claude write coding prompts for stages then I give the prompt to another agent. I wonder if I should write it up as a blog post
add another agent review, I ask Claude to send plan for review to Codex and fix critical and high issues, with complexity gating (no overcomplicated logic), run in a loop, then send to Gemini reviewer, then maybe final pass with Claude, once all C+H pass the sequence is done
Kiro's spec-based development looks identical.
https://kiro.dev/docs/specs/
It looks verbose but it defines the requirements based on your input, and when you approve it then it defines a design, and (again) when you approve it then it defines an implementation plan (a series of tasks.)
This separation of planning and execution resonates deeply with how I approach task management in general, not just coding.
The key insight here - that planning and execution should be distinct phases - applies to productivity tools too. I've been using www.dozy.site which takes a similar philosophy: it has smart calendar scheduling that automatically fills your empty time slots with planned tasks. The planning happens first (you define your tasks and projects), then the execution is automated (tasks get scheduled into your calendar gaps).
The parallel is interesting: just like you don't want Claude writing code before the plan is solid, you don't want to manually schedule tasks before you've properly planned what needs to be done. The separation prevents wasted effort and context switching.
The annotation cycle you describe (plan -> review -> annotate -> refine) is exactly how I work with my task lists too. Define the work, review it, adjust priorities and dependencies, then let the system handle the scheduling.
Pretty sure this entire comment is AI generated.
Almost think we're at the point on HN where we need a special [flag bot] link for those that meet a certain threshold and it alerts @dang or something to investigate them in more detail. The amount of bots on here has been increasing at an alarming rate.
There has been this really weird flood of new accounts lately that are making these kinds of bot comments with no clear purpose to making them. Maybe it comes from people experimenting with OpenClaw?
I don't see how this is 'radically different' given that Claude Code literally has a planning mode.
This is my workflow as well, with the big caveat that 80% of 'work' doesn't require substantive planning, we're making relatively straight forward changes.
Edit: there is nothing fundamentally different about 'annotating offline' in an MD vs in the CLI and iterating until the plan is clear. It's a UI choice.
Spec Driven Coding with AI is very well established, so working from a plan, or spec (they can be somewhat different) is not novel.
This is conventional CC use.
last i checked, you can't annotate inline with planning mode. you have to type a lot to explain precisely what needs to change, and then it re-presents you with a plan (which may or may not have changed something else).
i like the idea of having an actual document because you could actually compare the before and after versions if you wanted to confirm things changed as intended when you gave feedback
'Giving precise feedback on a plan' is literally annotating the plan.
It comes back to you with an update for verification.
You ask it to 'write the plan' as matter of good practice.
What the author is describing is conventional usage of claude code.
A plan is just a file you can edit and then tell CC to check your annotations
One thing for me has been the ability to iterate over plans - with a better visual of them as well as ability to annotate feedback about the plan.
https://github.com/backnotprop/plannotator Plannotator does this really effectively and natively through hooks
Wow, I've been needing this! The one issue I’ve had with terminals is reviewing plans, and desiring the ability to provide feedback on specific plan sections in a more organized way.
Really nice ui based on the demo.