alt Hacker NewsFrom my experience, a snippet of code will look good.
But LLMs seem to have a hard time getting the big picture and reusing code that is already implemented and ALMOST does what you want vs. rewriting everything from scratch.
AI code is extremely difficult to read and follow. It's littered with hundreds of lines of comments and notes, often referring to other notes in other parts of the codebase, and often extremely out-of-date.
Claude seems to me to still focus on getting things done vs. doing it properly, so from my experience it would pretty often mess up your whole codebase just to be able to finish its task, vs. stopping and rethinking the approach.
After a few passes of that, with duplicate code with no underlying coherent vision, hundred of thousands of lines of documentation written as walls of text in markdown and weird coding decisions, your codebase is impossible to work out for a human.
I use Claude Opus 5 at high to plan, Claude Opus 5 at low to execute the plan.
I make it build things in small changes, I give very specific instructions about the architecture, and I make sure to point out existing functionality that can be used instead of writing something from scratch. But it's a lost battle because Claude can't learn and it's basically a black box.
It pretty shitty, but that's what they pay me for so I just do it.
Replies
> It's littered with hundreds of lines of comments and notes, often referring to other notes in other parts of the codebase, and often extremely out-of-date.
Adjustable by tooling and prompts - my favorite is a ban on comments >2 lines, a tagged index system for effectively in-repo RAG with short documents on each tagged topic like INFRA-DATABASE-SETUP.md - enforce with tooling that the tag matches the doc and vice versa. Enforce that any PR that has a tag change in it has the relevant doc changes.
> Claude seems to me to still focus on getting things done vs. doing it properly, so from my experience it would pretty often mess up your whole codebase just to be able to finish its task, vs. stopping and rethinking the approach.
Tell it your values - "I value correctness over getting things done, and getting things done properly over speed" solves many of those problems. I've actually spent a portion of today rejiggering my values documents because the models have improved enough I no longer need to be quite so careful about scoping.
I was having those same problems, but particularly using Fable / Sol to do the meta-work has largely eliminated them over the last ~3-4 months. Codebase is now a lean mean token-conserving machine.
Thanks! Many of your issues sound more agent / LLM specific than LLMs in general. At least compared to my experiences.
It also aligns with what I saw from Claude & Claude Code when I used it last year for a while. Now I use Codex and don't see (nearly) as much of that sort of behavior.
> But LLMs seem to have a hard time getting the big picture and reusing code that is already implemented and ALMOST does what you want vs. rewriting everything from scratch.
Yeah that's probably the weakest point of LLMs still. GPT-5.6 Sol got much better at this for me. I still usually end up doing 2-3 iterations of prompts and exploration to clean up various things but usually it's pretty light work now.
> hundred of thousands of lines of documentation written as walls of text in markdown and weird coding decisions, your codebase is impossible to work out for a human.
Okay Claude definitely seems to have an issue there. From what I've seen from coworkers using Claude it generates reams of endless docs. I resist the urge to `rm docs/planning*.md`! I don't think they get how bad Claude is at that.
A month back I tried the latest DeekSeek and it made reams of text back and forth with itself, but the code it output was reasonable and it didn't make pages of markdown files either.
> I make it build things in small changes, I give very specific instructions about the architecture, and I make sure to point out existing functionality that can be used instead of writing something from scratch.
That's a bummer. I found that once the models start having problems that it cascades.
I've also been able to keep steady progress on a 70k+ LOC GUI side project without the endless whack-a-mole of bugs using Codex and Sol. Squash a bug, review architecture, move on, etc.