The code I check in at work is nearly all hand-written. I'll often send files and functions and snippets and questions and `git diff`'s to an LLM. In particular the code review is very helpful for catching silly mistakes. But if it generates code, even if it looks fine, I always end up manually moving it around and rewriting and renaming things. I need to understand the code I maintain, and those "prosaic" changes and refactorings that I do while incorporating the code help ground me deeply in how it works.
Also, I'm mostly not very happy with the architectural choices the LLM makes. It can be good at little details, and sometimes I learn about language features or idioms or libraries by reading its code, but I find it often uses wrong patterns. E.g. just this week I saw it creating two huge intermediate data structures for parsing something in a "cleaner" way, completely missing how the library it had suggested actually had something builtin (and prominently featured in the tutorial) that obviated the need for those intermediate structures and was both safer and more readable.
When I've tried full agent coding, I kept getting distracted while waiting. I also have the feeling that my questions to the LLM were getting "dumber" as I lost contact with the code. And I don't have the hardware for local LLM's, so giving full repo access is out of the question for work.
Altogether I feel like I've found a fairly good balance in how to use them in such a way that I can avoid a lot of the downsides and still have a tool that gives me much value that I didn't have before: I check in fewer bugs, I waste less time picking libraries/methods, I have a better rubber duck, I learn little tricks all the time, I never send LLM-generated text to humans.
Also, I'm mostly not very happy with the architectural choices the LLM makes. It can be good at little details, and sometimes I learn about language features or idioms or libraries by reading its code, but I find it often uses wrong patterns. E.g. just this week I saw it creating two huge intermediate data structures for parsing something in a "cleaner" way, completely missing how the library it had suggested actually had something builtin (and prominently featured in the tutorial) that obviated the need for those intermediate structures and was both safer and more readable.
When I've tried full agent coding, I kept getting distracted while waiting. I also have the feeling that my questions to the LLM were getting "dumber" as I lost contact with the code. And I don't have the hardware for local LLM's, so giving full repo access is out of the question for work.
Altogether I feel like I've found a fairly good balance in how to use them in such a way that I can avoid a lot of the downsides and still have a tool that gives me much value that I didn't have before: I check in fewer bugs, I waste less time picking libraries/methods, I have a better rubber duck, I learn little tricks all the time, I never send LLM-generated text to humans.