you'd be better off with two chats, one is your prompt creator, you pick the prompt, use another chat to create answer (zero shot chat cleared every time), and you copy the answer back into the prompt creator, with feedback on what the answer get wrong.

that's because from the point of view of gpt, gpt is always providing is most accurate anwer already, and the only improvement you'd get is in context clarity as defined by gpt understanding of language, which is not bad by itself, but it's not as effective as having gpt fight itself.