Ha, you just gave me an idea. Add to the prompt “do not do things that will burn over X tokens if the human operator can do it in less than X min, ask for it”.

I wonder if LLMs can estimate effort in tokens?