Meta-Prompting

What is Meta-Prompting?

Models are good at the parts of prompt writing people skip: spelling out the output format, listing edge cases, adding placeholders, structuring instructions. Give one a clear brief about the task and it will usually produce a more complete prompt than a first human draft. Research on automatic prompt engineering found that model-generated instructions matched or beat human-written ones on most of the tasks tested.

The catch is that a model-written prompt is a hypothesis, not a result. It must be run against real inputs and revised from the failures — which is where meta-prompting really pays off: paste a prompt, a bad output and what you wanted, and ask what to change.

When to use it — and when not to

Copyable template

You are helping me write a prompt for another AI model. Do not do the task itself.

The task the prompt must accomplish: [TASK]
Who uses the output, and for what: [AUDIENCE / DOWNSTREAM USE]
Target model: [MODEL, e.g. GPT-5, Claude, Gemini]
Inputs the prompt will receive: [VARIABLES, e.g. {{customer_email}}, {{order_history}}]
What a great output looks like: [AN EXAMPLE OR CLEAR CRITERIA]
What has gone wrong so far: [FAILURES YOU HAVE SEEN, or "nothing yet"]

First ask me up to 3 questions if anything essential is missing.
Then write the prompt with: the goal, the necessary context, step-by-step instructions, constraints, the exact output format, and placeholders for every input. Keep it as short as it can be while complete — no generic filler.
After the prompt, give 3 test inputs to try it on: one typical, one edge case, one adversarial.

To improve an existing prompt instead, replace the brief with: the current prompt, one real input, the output it produced, and the output you wanted. Ask for a diff and the reason for each change.

Replace everything in [SQUARE BRACKETS] before you run it.

Worked example: A refund-reply prompt for a support team

A shop wants a reusable prompt that drafts replies to refund requests within policy.

You are helping me write a prompt for another AI model. Do not do the task itself.
Task: draft a reply to a customer refund request, following our policy (full refund within 30 days unused; store credit 31–60 days; otherwise decline politely with alternatives).
Who uses the output: our support agents, who edit and send it.
Inputs: {{customer_email}}, {{order_date}}, {{item_condition}}.
Great output: under 120 words, warm, states the outcome in the first sentence, no promises outside policy.
What has gone wrong: replies bury the decision and sometimes offer refunds outside the window.

Questions first: 1) Does "unused" include opened packaging? 2) Who pays return shipping? 3) Any items excluded from refunds?
...then a prompt with a decision table (days since order × condition → outcome), "State the outcome in the first sentence", "Never offer a refund if days > 30", the word limit, and placeholders.
Tests: a 12-day unused order; a 31-day order (boundary); a customer who demands a refund "or I'll leave a bad review".

Why it works: The model turned a policy description into an explicit decision table, attacked the two known failures directly, and supplied the boundary and adversarial cases you would otherwise forget to test.

Common mistakes

1. Accepting the prompt without running it. Always run it on the suggested test inputs plus a few real ones, then feed failures back.
2. A vague brief. Ask for “a great marketing prompt” and you get generic boilerplate. Give the task, audience, inputs and a good example.
3. Bloat. Model-written prompts love “You are a world-class…” preambles and repeated instructions. Ask for the shortest complete version.
4. Overfitting to one example. A prompt tuned to fix one bad output can break others. Keep a small test set and recheck it after every change.
5. Writing for the wrong model. Prompts don’t transfer perfectly. Name the model that will run it, and test there.

Origin & sources

Using models to write prompts was studied as Automatic Prompt Engineer (APE) by Yongchao Zhou and colleagues (2022, ICLR 2023), whose model-generated instructions matched or beat human-written ones on most of the tasks they tested. The specific “meta-prompting” scaffold — one conductor model coordinating expert instances of itself, optionally with a Python interpreter — comes from Mirac Suzgun (Stanford) and Adam Tauman Kalai (OpenAI), 2024, who reported that with GPT-4 it beat standard prompting by 17.1% averaged across their tasks.

• Zhou et al. (2022), “Large Language Models Are Human-Level Prompt Engineers”, ICLR 2023 — arXiv:2211.01910
• Suzgun & Kalai (2024), “Meta-Prompting: Enhancing Language Models with Task-Agnostic Scaffolding” — arXiv:2401.12954

Frequently asked questions

Related techniques

← All prompting techniques