I still write code myself but I do use LLM. I use them for exploring ideas, explaining things, look up syntax, look up API/reference, look up algorithms, write tests, write duplicate code, write small functions that can be reviewed easily, do code reviews, find bugs, do performance review, do complexity analysis, etc.

The LLM generated code are correct about 80% of the time. Sometimes the generated code is overly complicated, has poor performance, missing a few assumptions, etc. that you need to review it carefully. It helps to have a solid and extensive set of tests.