It can be done, but for my environment the sum of all prompts that I end up typing to get the right result ends up being longer than the actual code.

So now I'm using LLMs as crapshoot machines for generating ideas which I then implement manually