I've actually been experimenting with minimal prompts; ideally, I'm familiar enough with the model's interpretation of the prompting structure that I can get good results without overdetermining the prompt to the point where there is no information beyond the narrow range allowed.

I think part of the defensible value proposition of these generic models is discovering your "unknown unknowns", and providing some long detailed prompt will eliminate that possibility; however, they're not as deterministic as programs that you write yourself for directly solving tasks, so they'll often fall short of the highest expectations, if you only use them as workers for solving completely-specified tasks, and even when they do perform some task perfectly, the total costs are ridiculous.