I've found that the only prompts that actually work for generating code reliably are the ones where you already know exactly what code it will output -- where n

I've found that the only prompts that actually work for generating code reliably are the ones where you already know exactly what code it will output -- where nearly every part of the "how" is precisely clear, it just needs to be compiled from human language into code. Writing this perfect prompt often turns out to be a waste of time. It's a classic specification problem. Human languages will let you under-specify a problem, but the result will not be pleasant.