[zerr]

Besides not wanting to give most enjoyable part of the work to LLMs, reviewing LLM-generated code is much more daunting compared to writing the code yourself. So, I only use it for a very narrow and specific 2-3 liners, e.g. some arcane Win32 API calls, where I'd otherwise be browsing some old forums.

[taariqlewis]

This is my reflection as well. I find myself spending MORE time reviewing LLM-generated code and also spending time thinking through LLM generated choices, which, at many times are inefficient or bloated. Keeping the LLM on the right rails takes up more time, even with lengthy agent.md and claude.md files to manage behaviors.