But the trick is that if / when you can define "good code" in a deterministic manner, then the LLM can also deliver "good code".

But if it's just based on feels, then of course it can't do it because it's not a mind reading machine.