Meta-prompting: ask the model to write the prompt

Meta-prompting means asking a model to write or refine a prompt and not to do the task itself. It earns its place for a prompt you will reuse and does nothing for a question you ask once.

The basic move

I want a prompt that does [task], for someone who will run it repeatedly with different inputs. Write that prompt, including what output format to request and what edge cases to guard against. The model drafts a structured prompt with the techniques a careful person would use: a named role, a specified output shape, a rule for missing or ambiguous input.

Where it helps

Use it for a task you understand well enough to describe and have not yet turned into a clean, reusable prompt. It beats a blank page. A model drafting a prompt also tends to include structural pieces, an explicit output format and a fallback for missing data, that are easy to forget when you write from scratch in a hurry.

Where it does not

For a one-off question it is pure overhead. Writing a prompt to write a prompt to answer one question takes longer than asking directly and adjusting. Keep it for something you will run more than a handful of times.

Iterating with the model

After the draft, paste in a case where the prompt gave a bad result and ask: here is the prompt, here is the input, here is the output it gave and here is what was wrong with it. Suggest a specific change to the prompt that would have prevented this. That is a far more targeted ask than make this prompt better, and you get a more targeted fix.

The caveat

The model is no more infallible at this than at anything else. Review a generated prompt as you would a colleague's, and test it on a few real inputs before you rely on it for anything that matters.