What is meta-prompting, and should you let the AI write your prompt?

A prompt is the instruction you give a language model. A meta-prompt is a prompt about prompts: you describe your task, and the model writes a detailed instruction that another conversation, or another model, will follow. The idea sounds circular, but it solves a real problem. Many people know what they want and still freeze in front of an empty box, unsure how much detail to give or how to structure it.

Researchers have studied this for a while. In a 2022 paper, Yongchao Zhou and six co-authors described Automatic Prompt Engineer (APE), a method in which a model proposes many candidate instructions and the one that scores best on test examples wins. On 24 language tasks, the chosen instructions matched or beat human-written ones on 19. That result does not mean a model always writes better prompts than you; it means a model can generate useful candidates, and testing picks the good ones.

What Anthropic's metaprompt does

In March, Anthropic published an experimental "metaprompt" for its Claude models, with a notebook in its public cookbook on GitHub. Running it requires an Anthropic API key, the access code developers use to call Claude from their own code. You type a task in plain words and, if you wish, name the inputs the prompt should accept, written in capitals, such as DOCUMENT or QUESTION. Claude then returns a full prompt template.

Behind the scenes, the metaprompt is itself a long prompt filled with about six examples of well-built prompts, from a customer-support agent that may only answer from a company FAQ to a patient math tutor. Seeing those examples, Claude writes something in the same style for your task: clear rules, placeholders for your inputs, and often a request to find evidence or think in tagged sections before answering.

Anthropic is candid about the limits. The notebook is designed for single questions and answers, not long conversations, and is tuned for Claude 3 Opus, the largest model. It also warns that the prompt you get is not guaranteed to be the best one and encourages you to edit it. The company presents it as a cure for the blank page and a quick way to produce several versions to test.

A worked example: meeting notes to action items

You don't need the notebook to try the idea. Any capable chatbot can play the same role. Suppose you want a reusable prompt that turns rough meeting notes into a list of action items.

Ask for the prompt:

I want a reusable prompt that turns rough meeting notes into a list of action items. The notes are typed quickly, with abbreviations and missing words. Each action item needs an owner, a task and a deadline if one was mentioned. Write the prompt for me. Use {NOTES} where the notes will go. Include rules for what to do when the owner or deadline is unclear.

A reply might look like this (illustrative):

You will receive informal meeting notes in <notes> tags.
<notes>{NOTES}</notes>
1. Read the notes and list every commitment someone made.
2. For each, write: Owner | Task (start with a verb) | Deadline.
3. If the owner is unclear, write "Owner: ?" and quote the line it came from.
4. If no deadline was stated, write "none stated". Never invent a date.
5. Do not include discussion points that were not commitments.
Return only the table.

Your own version, written by hand in two minutes, might be: "Pull the action items from these notes as a table with owner, task and deadline." It is shorter and perfectly reasonable.

Now test both on the same three sets of notes: one tidy, one messy, one with no clear owners. The results below are illustrative, but they show the kind of difference to look for:

The AI-written version wins here because it anticipated the awkward cases. That is the main strength of meta-prompting: the model thinks of rules you might not have written.

Should you let the AI write it?

For one-off questions, usually not; a plain request and a follow-up or two will be faster. For a task you will repeat, or a prompt you will share with colleagues or build into a tool, a generated draft is a good start. The risk is that a long, confident prompt looks finished. It can include rules you don't want, assume facts about your situation that aren't true, or be tuned for a different model than the one you use.

How do you check the prompt it writes?

Read it line by line first. For every rule, ask whether you actually want it. Delete anything that doesn't match your situation, and add what the model could not know, such as your audience, your house style or the length you need. Make sure every placeholder is one you will really fill.

Then test it the way APE does: on examples, not on impressions. Gather three to five real inputs, including at least one awkward case, and run the prompt on each. Write down in a sentence what a good answer looks like before you look at the results, so you judge against a standard rather than whatever appears. Compare against your own simpler version; if the short prompt does as well, keep the short one.

Finally, retest when something changes. A prompt that works with one model may behave differently with another, and Anthropic itself notes that its metaprompt was built for one specific model. Keep the tested version and its test examples together, so you can check it again in a few minutes.