I've found they optimize largely for the happy path and don't consider any (or enough) edge cases (e.g. what happens to everything downstream of this, if say, there's a timeout in this specific HTTP request.)
I find Claude fighting me when I point out things like that as well claiming its not worth it to worry about and when I point out the train wreck that would ensue by leaving some things in a weird state it flips its script: "I was incorrect about that..."
Folks will claim that's a skill issue but, its an issue letting the LLM run off without any oversight and it creates so much code you can't possibly review it all yourself, so like you've found, you hit a lot of problems in production.
AGENTS.md and friends help but I've found it ignoring rules in their as well, often and very frequently.
Personally what I've found to be the biggest win is:
1. Use AI to go harder not faster - use it for code reviews, second opinions, researching all the angles to a deep topic, but don't use it to pump out a whole app. Don't outsource your entire understanding to it (which is precisely what we do at my job today...)
2. Use it for the boring tasks that are hard to get wrong and can be easily validated.
Oh and tests... I've seen so many completely useless tests being generated. Any valuable test that I've ever seen claude create came from me finding a bug or missed edgecase.
This is what I'm seeing at my company as well, "software development principles" are out the window. People have forgotten all about "Second System Syndrome" too, docs, tests, code, design - all LLM generated.
I don't like, I prefer to use Claude/Codex as an aide, not to let it take the wheel entirely but I don't really feel like I have much choice in the matter.
We get scolded for asking other people on our team questions sometimes, and get told "you should ask claude that", I don't even communicate with my team anymore, instead of asking someone with domain expertise about something specific I just ask claude now... I have no interest in my work whatsoever anymore, it used to be what I did night and day, happily.
Non-engineering higher ups use claude to do system design, use claude to write docs, throw together an MVP that is barely and often not at all functional. Then we have to actually make the mess work... of course, with more claude. I hate everything about it.
I find Claude fighting me when I point out things like that as well claiming its not worth it to worry about and when I point out the train wreck that would ensue by leaving some things in a weird state it flips its script: "I was incorrect about that..."
Folks will claim that's a skill issue but, its an issue letting the LLM run off without any oversight and it creates so much code you can't possibly review it all yourself, so like you've found, you hit a lot of problems in production.
AGENTS.md and friends help but I've found it ignoring rules in their as well, often and very frequently.
Personally what I've found to be the biggest win is:
1. Use AI to go harder not faster - use it for code reviews, second opinions, researching all the angles to a deep topic, but don't use it to pump out a whole app. Don't outsource your entire understanding to it (which is precisely what we do at my job today...)
2. Use it for the boring tasks that are hard to get wrong and can be easily validated.
Oh and tests... I've seen so many completely useless tests being generated. Any valuable test that I've ever seen claude create came from me finding a bug or missed edgecase.