I completely agree that the best LLM output is something that is still understandable/hackable -- and maybe even _needs_ further hacking. That's what I tell people when they say "I heard AI is replacing programmers."