Depends on the language. Recently I got stuck on some issues with line-based processing of stream input with Haskell conduits and resorted to ChatGPT out of desperation. It wasn't of much use to me either, and then I ran out of free tier usage.

I suppose for emitting boilerplate or other ceremonial code LLMs serve as an augmented "copy/paste from stackoverflow" and increase your KLOC/time, but this is not really an interesting metric and production of source code was never the rate-limiting step in creating software.

You used the wrong language. Switch to one that has enough LLM training data :)

Read data into x?

Taking into account that the file may not exist or error half way?

Taking care that your buffer doesn’t overflow or simply eat all your available ram?

Do you only work with trusted inputs?

Most the time, yeah, actually.

Also- in my experience so far, LLMs do a far better job of accounting for things like this than I do.

For example I had claude 4 write something to recursive search the whole filesystem on a windows 11 machine and included a few checks for windows 11 file system specific things I wasnt aware of