I think most of this is correct, in spite of potentially being built on a bad assumption.
The assumption is that LLMs should be writing the code and human engineers reviewing and verifying the LLM output. And that this pushes the cost of producing down. And I fundamentally disagree with that.
Every time I ask LLMs to write code, even with Opus 4.8 (haven't tried it with Opus 5 yet), what I get ends up being totally rewritten. LLMs still aren't good at writing maintainable code. Can they write plausibly functional code? Yes. But it won't survive the long term. People using LLMs to write all their code are gambling on them eventually getting to a point where the LLMs can fix their own code. It's possible, but I wouldn't necessarily bet on it.
Where I have found immense value from LLMs is in code review. Repeated review by LLMs catches an amazing amount of potential issues. They really shine on security review, but are very effective with any kind of review.
The other thing that the "LLMs write code camp" misunderstands is that writing was never the bottleneck. Understanding was. And understanding the code is still the bottleneck. But understanding is truly gained during the writing loop. The understanding you gain from pure reading or code review is marginal compared to the understanding you gain while writing.
Most of the time previously spent writing was actually spent updating and deepening our understanding of the system under development. There's no replacement for that understanding in a world where LLMs are doing the writing.
But if you flip it: humans write, LLMs review, then you still get a major gain -- not in speed, but in quality. And you keep the understanding loop intact. I would propose that this might be the best way to deploy LLMs.
I think this is outdated. If you follow spec-driven development, get the model to do all the planning work upfront, review and iterate the plan, write clear markdown file documentation on the abstractions and patterns you want to follow, then you have every opportunity to tell the model how you want it to write the code. If you use Opus or Fable 5 it will then write the code better and faster than you will.
you're a lot more skilled than I am if you're able to know the abstractions and patterns you want while only have a weak grasp on the actual code. For me the abstractions emerge as I understand what I'm actually working with
I think you can have some system design in mind upfront, but agree the bulk emerges as you code. Clean Code practices worked exactly this way in which you write and then rewrite the clean code.
I agree, but this works too because AI makes prototyping more rapid. It's common to be stuck with suboptimal solutions because of time constraints. Maybe the feature isn't really what the customer wants, or maybe the code isn't actually very good. But it reaches prod anyway.
But, with AI, the cost of code has gone down more than it already has. Well, cheap things are easy to throw away. So you prototype, prototype, prototype, and close the loop as much as possible with the customer. True agile development, not big A Agile.
The problem is this requires alignment from management, and we're just not seeing it at many company. They can't grasp that things have changed, and that throwing away code is free. They don't trust engineers to close that gap, so customers and stakeholders are still waaaaaay over there and we're delivering features they don't want.
IMO "get the model to do all the planning work upfront, review and iterate the plan" is backwards. It works better if you do an initial plan yourself and then have the AI review it (and iterate as necessary). Makes you think about the design for a bit so you can have a semblance of a mental model.
I have never seen a plan written by LLM that wasn’t vague and light on details, every time I need to ask for more details and every time I ask to implement it trips over some dead end in the plan, discards the plan and continues as if there was no plan to begin with. Planning feels often just narrating the request and a wish list then what a human would do - methodically build enough understanding so that you’re confident of the direction you choose. LLM plans are overhyped and overrated.
This just doesn't match my experience writing idiomatic C code for embedded real-time systems with ChatGPT-5.5 and -5.6. Are Opus and Fable really that different? I general I find that a fully correct implementation for some function may take an hour and be nearly one-shot, but going from there to code I'd be willing to actually commit is at least 4x the time commitment with regular back-and-forth -- so four hours of deep engagement for every one hour of one-shot. ChatGPT-5.6 Sol seems to be better that -5.5 at the one hour do-it portion, but worse at the four hour do-it-well portion.
Sure. But try it on a project on non-trivial size and then tell us the cost of it. LLMS are billed per token.
Depends on how long that planning/review cycle takes.
True, but I think people are also missing the trick behind markdowns that reference each other and are used for context window management. They go in your source control, they are reusable, extendable, and composable in the same way code is. So planning is not a one-off effort. It compounds over time. If you're doing it right your plan is referencing automation techniques - testing, CI/CD, etc. This further compounds so that the agents verify their own work against the standards that you set.
To be honest, the models are getting so good that they do most of this unprompted now.
I haven't done "spec driven development" in my entire career, 15+ years
Until managers give up on Agile, we'll never get time to actually write specs
Agreed, all of the staff/principal engineers I work with are ~99%+ AI generated code, and increasing business value delivered as a result. This is on planet scale infra not CRUD apps. (And yes, you do need to carefully review the output and give steers/corrections. It’s still faster.)
At this point if you can’t get the agent to write good code then either I) you are in a very specific niche (like Karpathy trying to write NanoGPT) that is extremely out-of-distribution, or II) skill issue, you need to learn how to prompt better.
It’s fine to have a skill gap! Just don’t delude yourself that the tools are bad and everyone claiming they are good is wrong.
How is a GPT implementation “extremely out of distribution” but “planet scale infra” isn’t? That’s got me totally confused about your point that I was taking seriously.
A lot of stuff that underpins planet-scale infra is open source and has been for a long time. The general approaches and architectures for it have also been widely discussed and litigated in the commons so it's not just the code - the specific why is also very well documented.
So despite its importance much of it is actually pretty in-distribution.
Okay, but then that does not square with calling a GPT implementation out-of-distribution, let alone "extremely" so.
it's really cool how in the past measuring developer productivity was incredibly difficult, but now its become easy if someone's making a point about LLMs
btw, your last paragraph makes you sound like an ass.
> writing was never the bottleneck. Understanding was.
Well put. This insight is worth repeating in every discussion on the subject, from software engineering to mathematics.
The problem is that understanding is not the product being sold. The business model is for everyone to become consumers of what the magical genie generates, where the "understanding" is kept on the side of the model providers. This ensures a future generation of consumers dependent on someone else to provide the understanding.
Otherwise, you can create your own answers based on actually understanding the code, theorem, proofs, etc. Smart consumers of LLMs will use them to increase their own knowledge and understanding, so that the service is augmenting their intelligence, not replacing it.
How do you handle the fact that an LLM can't seem to help itself from disgorging page after page of words no matter what it's asked to do? I've never seen an LLM say "this looks good as-is; I would not spend any more time on it; what's next?" it will always seem to suggest using another pattern or additional abstractions or other yak shaving.
But to be fair human code reviews have the same problem. It's like reviewers feel they have not done their job if they don't find something wrong.
I've totally had latest models just say "here are minor nits but this is good to ship" when reviewing code. Claude is a bit more verbose usually but Codex by default is pretty terse. My experience anyways. They also take pushback on suggestions (this isn't actually a problem because X) and they'll agree and say ship it (hopefully only when you're right and it's not a problem ;) ).
I personally rely on this behavior of LLMs. For example, for reviewing prose, I might ask the LLM to find five issues. If there are obvious issues I missed, then it'll find them. But if all the issues it lists are minor or hallucinated, then I can have some confidence there weren't any glaring issues.
I do the same for LLM code review comments: some changes are out of scope or could be moved to a separate PR; some edge cases don't happen in practice and should just fail noisily instead of writing more code to maintain. When these are the only issues it's raising, then I know it's done.
I've tried and tried to get an LLM to delete code, it couldn't do it. I knew a file was 40% bad so what I ended up doing was deleting the file, then asking AI to create the missing file. That's how I got AI to delete code :)
Sounds like the AI got you to delete the code for it.
It is surprisingly hard to do in a single prompt. I’ve had good luck though with being very explicit about asking for review, then to think deeply about what actually matters, and then reduce to a very short extremely concise reply. —- the model NEEDS to output those pages of text as part of its thinking process. It’s used to doing it in the assistant reply but can be cajoled into doing it in the thinking tokens.
Yeah, that's just a problem with review. I tend to keep it going until the stuff its finding is stuff that I look at and think "I can live with that". Usually by that point it's found the worthwhile stuff and what remains is minor or, like you said, yak shaving.
There used to be a bit of "wisdom" passed around that said "when you submit something to the client (or management, etc) for approval, leave an obvious mistake somewhere." Then they will find it and point it out, you can easily correct it, and everyone is happy. Otherwise they will find something to critique, just to demonstrate that they contributed something to the review.
I wonder if the same trick works with AI?
I think so, if you can inject synthetic errors with the same distribution as the errors you care about.
What tools are you using? I've found that Codex will happily tell me there's nothing more to do after a few rounds of /review and fix (or sometimes, it picks up on low value points that are technically correct but not worth changing).
My guess would be that this is a matter of fine-tuning the model to the problem. You would need to run the review experiment N times, observe the distribution of answers, and then fine tune with model operational parameters.
Perhaps someone more knowledgable could jump in here to clarify?
I’ve found stating “only call out critical issues, report but ignore everything else” will basically get what you want. Put that in a review skill and you’re good to go.
To be clear I mostly only use Opus and Gemini Flash but this might work for others too.
> I’ve found stating “only call out critical issues, report but ignore everything else” will basically get what you want.
Disabling your C compiler warnings works too. You get to ship then leave work early!
The verbosity itself seems to be a problem with post-4.6 Claude rather than a general issue with all LLMs. and IME the yak shaving is due to an overly generic prompt. We have a generic automated review bot but it’s prompted to only look for errors and never suggests refactoring; that was a conscious trade off to avoid what you’re talking about. If you’re reviewing interactively then you can give additional guidance about any code smells etc. that jump out
I’ve found that doing the full requirements capture, planning, writing, reviewing, gardening loop with frontier models has worked quite well since last October, and phenomenally since Fable 5.
The key I’ve found is human peer review. The reviewer jumps on a live call with the developer, pulls up the PR with transcription on, and asks questions. At the end of the call, the transcript passes back into the coding agent and the PR is polished up, becoming more self-documenting, and the humans are left with some degree of common understanding of what’s going on.
I’ve been operating my team of ~15 this way for 9mo to great effect… there is simply no going back to the stone ages.
Here's an odd request for you, as a fellow New Jersey-an...
Can I watch/observe one of your review sessions?
Every few weeks, I hear the beginnings of a great approach towards working with LLMs but I rarely see it in practice.
If you're open to this, remote or in person, ping my username at gmail.
I have not experienced this "llms dont write maintainable code narrative" I just tell it the shape of the entities and apis I want vaguely and the mental model and the output is excellent.
I could do this in the beginning of 2025. Unfortunately, given the over-proactiveness of model these days, I find that they end up inferring what my original request was and implementing it anyway, and then writing unit tests, etc. even when all I asked for was to wire up a few components, as an example.
I actually find the converse, that I try to scope things down into chunks small enough for the model to work on, but it will start trying to hack off pieces that don't belong because it thinks the restricted scope is the whole world and I have to stop it and explain the bigger picture so it stops trying to remove things that serve the full scope.
I write code in Go in an established codebase. 90% of the time the LLM gets a basic api/feature right and fully e2e tested as long as I give it enough business context in the prompt. The other 10% of the time I have to do some follow up prompts to either change the behaviour or change the approach of a given step in the flow or just to point out that the wrong pattern was used and please rather use our codebase standard. I haven’t actually hand-written code in well over a year. Is this way faster? Absolutely. Does it lead to better code? Yes, because I have time to write all the regression tests that keep the behaviour as I wanted it when others come bungling around.
I do think the comment is right about understanding. I'm still feeling out what I think is the right level of understanding to invest in now. I think it is probably not the deep line by line understanding I used to have of the systems I worked on. But I also think it's easy to remain too aloof from how the system is being built, and that this is very bad. I think the right answer is somewhere in between, but I'm still working through my own process and forcing functions to strike that balance properly.
> The other thing that the "LLMs write code camp" misunderstands is that writing was never the bottleneck. Understanding was. And understanding the code is still the bottleneck. But understanding is truly gained during the writing loop. The understanding you gain from pure reading or code review is marginal compared to the understanding you gain while writing.
This was my stance a couple of years ago, but now I've given it up.
It turns out writing actually was the bottleneck. You can understand perfectly well what you want, but writing it is long and tedious to the point where you find excuses not to do it. Particularly with version 2, the step where you have an OK system and you want to improve it. Quite a lot of changing the code is just useless busywork: re-wiring old functions, moving imports around, searching for locations that benefit from extracting a common piece of code. And each time you do one of those, there's a decent chance you did something even more trivial like forgetting a semicolon or calling the wrong function.
Now that I have an LLM helping me, I can see why. The critical decision is a terse declarative like "we need to have several TCP connections instead of one, and just use the sequence number to arbitrate". A human junior programmer could perfectly well understand what this meant, but he would have to go through all of the above to get to the final product. Now, I can just tell the LLM and I will get what I want, even with the things I didn't explicitly state, without spending attention.
This means I can use my attention on the things that matter. So instead of spending today thinking about how to arbitrate between the TCP connections and tomorrow thinking about pre-calculating my outgoing orders, I can just do both today. I don't waste the good waking hours chasing minor bugs, I just think about the large structure.
I get the feeling the best programmers of years past were actually masters of the little things, which led them to be able to look at the big things. Essentially it was cheaper for them to get to the top of the mountain, where you can see the landscape. Kinda like how the kid who was good at mental arithmetic in primary school was also good at calculus at the end of high school: if you don't have to concentrate on the little things, you have time for the big things.
> And each time you do one of those, there's a decent chance you did something even more trivial like forgetting a semicolon or calling the wrong function.
how did you decide to pick the most trivial kind regression for this example? do you compile your code before checking it in?
> A human junior programmer could perfectly well understand what this meant, but he would have to go through all of the above to get to the final product. Now, I can just tell the LLM and I will get what I want, even with the things I didn't explicitly state, without spending attention.
the main efficiency you have described here is offloading the verification of a change onto the LLM. that is the bottleneck. readers can decide whether a non-deterministic statistical model is a good tool for this job
> I get the feeling the best programmers of years past were actually masters of the little things, which led them to be able to look at the big things
the best programmers understand that their job is to automate workflows, and that includes their own. if you're worried about missing a semicolon, I'm sorry to say that's a skill issue
> how did you decide to pick the most trivial kind regression for this example?
Why would this be a regression? You might just be writing a new line of code.
> do you compile your code before checking it in?
Well obviously. That is generally how you discover that a semicolon is missing.
> the main efficiency you have described here is offloading the verification of a change onto the LLM. that is the bottleneck. readers can decide whether a non-deterministic statistical model is a good tool for this job
No, it's the time between you deciding something needs to be done, and it being done, that is the bottleneck. You cannot avoid trying to compile the code and testing it. Now you can get to that test without paying attention, which is time you can use productively.
> readers can decide whether a non-deterministic statistical model is a good tool for this job
Somehow, the non-deterministic model has built me the deterministic code that I want, very fast, pretty much all the time. A year ago it would get stuck. Now it doesn't, for me at least, and for competent programmers that I know.
> the best programmers understand that their job is to automate workflows, and that includes their own. if you're worried about missing a semicolon, I'm sorry to say that's a skill issue
Well yeah, and I've automated my workflows completely. I don't have the problems I used to have. If you haven't caught on to the new way of working, well, that's a skill issue...
> Somehow, the non-deterministic model has built me the deterministic code that I want, very fast, pretty much all the time. A year ago it would get stuck. Now it doesn't, for me at least, and for competent programmers that I know.
I still find the models get stuck or go on _massive_ side quests. Just today, I asked claude to write a hello world C++ program using import std; I interrupted it when It decided I needed a new toolchain installed, and started checking for docker installations. This is super basic stuff, it hadn't even generated a plan, it just started searching for LLVM versions rather than running clang --version.
> If you haven't caught on to the new way of working, well, that's a skill issue...
Honestly, it feels like the emperor has no clothes on this topic, and the crowd defending LLMs to death are way too quick to call it a skill issue.
> Honestly, it feels like the emperor has no clothes on this topic, and the crowd defending LLMs to death are way too quick to call it a skill issue.
I feel it's the other way around. The LLM skeptics are unwilling to admit that these things can get you there faster than you would on your own, in the face of clear evidence.
> Well yeah, and I've automated my workflows completely. I don't have the problems I used to have. If you haven't caught on to the new way of working, well, that's a skill issue...
I use LLMs, but they're just a tool in the workflow, and I make sure to review the output. they might remember semicolons but they make much more pernicious mistakes that are harder to detect
But are those pernicious issues more common than what you would write yourself? Keeping in mind that you will get your LLM code back a lot sooner and have more time left over to check for them?
> Quite a lot of changing the code is just useless busywork: re-wiring old functions, moving imports around, searching for locations that benefit from extracting a common piece of code.
If other people are dissatisfied with LLM output quality while it seems to work fine for you, you might want to consider that the quality of code you produce is closer to the quality of code the LLM produces than what those other people are producing.
What you posted there, for example, about most of changing code being busy work is a pretty big red flag for a codebase. One of those "large structure" things that you're supposed to be paying attention to is the architecture of the code. There's always the chance that some change you need to do goes against the grain of the solution you architected, and you need to make changes all across your codebase to fit it in, but in general the point of modularity and good architecture is that when you make a change you just have to make that one change, ideally just changing the logic of the one responsible function with only minor changes required anywhere else in the codebase. If you're consistently having to hunt throughout the code for related functions that you need to rewire that's a sign that your architecture does not fit with the direction your codebase is evolving, or alternatively that you don't have much of an architecture to begin with and your code is highly interconnected.
Actually one habit you mention at the end of that quote can worsen this issue: "searching for locations that benefit from extracting a common piece of code". Tautologically this is a good thing as you define it as only working on locations that will benefit, but given the frequent need for rewiring of functions I would hazard to guess that you've "deduplicated" code a bit overzealously. Just because two functions share some common code does not necessarily mean it is appropriate to pull that out into a function. Deduplicating is good if conceptually the code is a single thing that you would always want to keep in sync, as it means that when you need to make a change to it you don't have to hunt down all the places it's used. On the other hand, if you find yourself frequently needing to delve in to these functions to rework them because you need to make a change to how it's used by just one caller, your "deduplication" has added to your workload, and probably created some overcomplicated code in the function that is in reality handling multiple distinct needs.
I hope this doesn't come across as too condescending, and if I've just wasted your time explaining principles you already understand I apologize. I don't know you or the code you're working on so I can't exactly confidently judge your work solely on a few paragraphs. It's just that your mention of how your experience of coding has been different from what others have described, and specifically that, for you, writing has been the bottleneck rather than understanding, combined with the specific issues you describe facing, imply to me that you may not realize that the approach you are taking to producing code yourself may be significantly different from how other Software Engineers are producing code, and that may account for some of the differences you note in your personal experiences programming.
I understand what you mean. Those are just some small examples. But I had this conversation with a friend today, and it goes something like this:
1) It was good for me to spend years learning the little stuff. Loops, variables, if conditions, how to import stuff, git, debugging things, reasoning about the flow of control. Classic coding.
2) I had a false dawn at about 10 years in. I thought I understood a lot.
3) I learned I had a lot to learn. Very wide areas of programming I'd never touched, ways of thinking that started to click.
4) I spent another ten years covering holes, building a different type of experience. My guesses about how to do a project are much better now. My guesses about what really matters have changed.
5) Now the small stuff is actually just bothering me. I'm not going to learn much more from staring at little things. There are larger architectural things to think about, and the little things are just friction.
So that's where I'm coming from. I get that a lot of pushback is going to be from 10-year-me, who thought he'd gotten to a high level of understanding by slogging through the little stuff.
>It turns out writing actually was the bottleneck. You can understand perfectly well what you want, but writing it is long and tedious to the point where you find excuses not to do it.
Speak for yourself. Writing code has never been a bottleneck for some of us. I can't speak for everyone, and neither should you.
>Now, I can just tell the LLM and I will get what I want, even with the things I didn't explicitly state, without spending attention.
This should worry you. All too often the LLM invents things I didn't ask for and implements things I didn't need. YMMV, I guess. If slop gets the job done, and nobody notices, then who should care?
This is where I've landed too, I imagine based off all the responses that it depends on domain and also the rigor you care about in your implementation. For me, what I do, and the quality required, there's just not much time (if any) saved on handing off the implementation to an LLM.
But that doesn't mean they're useless either. I use them all the time for review as you mentioned or to knock out one-off scripts that don't go anywhere near source control. There's just no world where I don't need to understand every line of code that I'm responsible for getting into our project.
To be honest, it feels like you are contradicting yourself.
If LLMs are, as you stated, really good at catching potential issues, then they are, almost by definition, really good at producing code without potential issues, if guided correctly: all they need to do is inspect and iterate, until they do not find any more potential issue in the code they produced.
That isn’t necessarily true. LLMs could be good at finding issues that human miss when coding, while being bad at catching issues that they themselves miss while generating code.
In my domain since circa Opus 4.5 LLMs write code as good as most engineers given well defined small enough chunk of work. They refactor. They write tests. These days LLM also debug/troubleshoot better than most engineers.
Are they as good as handcrafted code by 0.1% of top software engineers. Generally no. But neither is 99.9% of real code.
LLMs also are good at code reviews. What they'll miss is often the big picture but they can still catch plenty of issues. I still want to see a human in the loop in my domain.
Totally agree that writing the code was never the bottleneck. We're not seeing massive productivity gains even if some code is written faster. It's not just about understanding but also various other activities that happen in large companies and teams.
Also agree LLMs can be used to gain quality but realistically most orgs are going to aim for "fixed or decreasing" quality at lower costs.
There is the flip side of where the code is no longer read by humans and is becoming the prevailing way software is shipped in tiny businesses. You don't need to code to be maintainable since you will never maintain it, the AI will and the quality will naturally improve as models improve. For example 5.6 sol and fable are showing signs where you can feed garbage in and it will spit out something pretty decent, definitely not the quality you'd expect from a senior developer with millenia of experience, but that of your average grunt worker turning words in an issue board into code.
However, sometimes then I tell it to write an app with detailed instructions and it spits out garbage so your mileage might vary.
" But understanding is truly gained during the writing loop. The understanding you gain from pure reading or code review is marginal compared to the understanding you gain while writing."
Debugging code step by step is how I understand complicated code.
“Not in speed, but in quality” — unfortunately though, not a single CEO, executive, company, VC, investor or anyone with the power to make decisions cares about quality instead of speed.
I’m happy to report that I work for ones that do. Privately owned company with decades-long history, no VC, no profit maximization mindset.
I agree. Also, what does "code quality" even mean in the age of agentic dev? If the code works, and is secure, what else matters?
I do know what good code looks like, but does that even matter anymore? All I know is that now, I get to focus on endless UX polish, which is the only thing the matters.
I feel like we are living through something like the Protestant Reformation, where priests once spoke Latin, and then started to speak in plain local language. The old guard did not like this.
You're right about understanding. Where we might disagree is the conclusion. I think it's possible to build the understanding without having to type out syntax by hand (debugging, writing tests, etc). Maybe we're missing much better verification tools? LLMs will likely play a big part in those too (explain the codebase, walk me through A, B, C etc)
I've come to realize that the smarter LLMs get, the less they understand the point of abstractions.
I've been using it for a unity game for the past few years. Nowadays it will go sleuthing into packages and assembly and make decisions based upon what it sees there.
It will make comments about why it's doing something based upon a function call 3 methods deep.
God forbid any of these details change in a minor version update.
Try reducing thinking and use lighter models. For example I observed that using Sonnet works much better (compared to Opus) for tasks I want to be in control of architecture and just need a faster code input.
many of us work with swarms of cheap off-shore contractors and have to rewrite or finish their code already as it is. The llm is much lower friction
IME, the LLM when used well also gets closer to correct.
I hate saying this but unironically opus 5/gpt 5.6 are writing maintainable code on the order of 500-1000 lines
> LLMs still aren't good at writing maintainable code.
I find that depends on the target language. They can be good at writing maintainable code, but not consistently across every language.
The languages beginners usually gravitate towards are especially hard for LLMs to produce quality output for. Presumably this is due to the training data including all the unmaintainable codebases written by beginners in those languages, which hasn't allowed the LLM to converge on recognizing what a maintainable codebase looks like in those languages.