alt Hacker NewsIn my experience (with ChatGPT 5.1 as of late) is that the AI follows a problem->solution internal logic and doesn't think and try to structure its code.
If you ask for an endpoint to a CRUD API, it'll make one. If you ask for 5, it'll repeat the same code 5 times and modify it for the use case.
A dev wouldn't do this, they would try to figure out the common parts of code, pull them out into helpers, and try to make as little duplicated code as possible.
I feel like the AI has a strong bias towards adding things, and not removing them. The most obviously wrong thing is with CSS - when I try to do some styling, it gets 90% of the way there, but there's almost always something that's not quite right.
Then I tell the AI to fix a style, since that div is getting clipped or not correctly centered etc.
It almost always keeps adding properties, and after 2-3 tries and an incredibly bloated style, I delete the thing and take a step back and think logically about how to properly lay this out with flexbox.
Replies
> If you ask for an endpoint to a CRUD API, it'll make one. If you ask for 5, it'll repeat the same code 5 times and modify it for the use case. > >A dev wouldn't do this, they would try to figure out the common parts of code, pull them out into helpers, and try to make as little duplicated code as possible. > >I feel like the AI has a strong bias towards adding things, and not removing them.
I suspect this is because an LLM doesn't build a mental model of the code base like a dev does. It can decide to look at certain files, and maybe you can improve this by putting a broad architecture overview of a system in an agents.md file, I don't have much experience with that.
But for now, I'm finding it most useful still think in terms of code architecture, and give it small steps that are part of that architecture, and then iterate based on your own review of AI generated code. I don't have the confidence in it to just let some agent plan, and then run for tens of minutes or even hours building out a feature. I want to be in the loop earlier to set the direction.
Bad code is only really bad if it needs to be maintained.
If your AI reliably generates working code from a detailed prompt, the prompt is now the source that needs to be maintained. There is no important reason to even look at the generated code
> If you ask for an endpoint to a CRUD API, it'll make one. If you ask for 5, it'll repeat the same code 5 times and modify it for the use case.
I don’t think this is an inherent issue to the technology. Duplicate code detectors have been around for ages. Given an AI agent a tool which calls one, and ask it to reduce duplication, it will start refactoring.
Of course, there is a risk of going too far in the other direction-refactorings which technically reduce duplication but which have unacceptable costs (you can be too DRY). But some possible solutions: (a) ask it to judge if the refactoring is worth it or not - if it judges no, just ignore the duplication and move on; (b) get a human to review the decision in (a); (c) if AI repeatedly makes wrong decision (according to human), prompt engineering, or maybe even just some hardcoded heuristics