> LLMs don't write code like that. Many people have tried with many prompts. It's mostly good for just generating stuff once and maybe do little modifications while convincingly arguing the code has no bugs (and it has).

You have to make them write code like that. I TDD by telling the LLM what I want a test for, verify it is red, then ask it to implement. I ask it to write another test for an edge case that isn’t covered, and it will.

Not for 100% of the code, but for production code for sure. And it certainly speeds me up. Especially in dreadful refactoring where I just need to flip from one data structure to another, where my IDE can’t help much.