What it is
Having AI write the prompt, sometimes called meta-prompting, means asking a model for the prompt instead of the answer. You describe the task, and the model returns a structured brief: sections for the material, the task, the format and the rules, laid out the way a careful prompt writer would.
The model is good at that structure. It knows a prompt should say who the output is for, what shape it takes and what to do when something is missing. What it does not have is the content of those sections. Ask for “a prompt for summarizing sales calls” and it will decide what your managers care about, how long summaries should be and what counts as a good call. Those decisions sound reasonable and may all be wrong for you.
So the useful version is a collaboration. You supply the facts and the standard. The model supplies the structure, leaves labeled blanks for what changes each time, and asks about anything it would otherwise invent.
When it helps
Use it on a blank page, when you know the task but not how to brief it. Use it to tidy a rough prompt into a clear layout, or to adapt a prompt you wrote for one assistant to another, where the vendors’ layout advice differs. Use it to improve a prompt that gives inconsistent results, with the improve a prompt template. And use it to draft longer instructions, such as a system prompt or a Custom GPT’s instructions, that you would otherwise write from scratch.
If you already had a good conversation about the task, there is a better source: turn the chat into a prompt, so the rules come from your real corrections rather than the model’s guesses.
When it does not
If you do not have the facts, the model fills them in. A generated prompt that says “summaries should cover pain points, budget and next steps” reads like a decision someone made. Nobody did.
Generated prompts tend to be long and emphatic. They often add a role line, a list of rules that repeat each other, capitals for emphasis and “think step by step”. Current vendor guidance points the other way: discussing tool instructions, Anthropic advises softening aggressive language such as “CRITICAL: You MUST”, and OpenAI reports that leaner prompts, with repeated instructions removed, can improve results while using fewer tokens. Cut what does not carry a fact or a rule.
A prompt that reads well is not a prompt that works. Run it on real inputs before anyone relies on it, especially awkward ones. And if a prompt already works, be careful with “improve this”: a rewrite can quietly drop a rule. Ask for every change to be marked.
For a two-line request, writing the prompt yourself is faster than reviewing one.
Before and after
Two requests for a prompt, from different kinds of work. The scores are the Prompt Grader’s free estimate with Written for set to Any chat model. A score counts what the prompt states; it does not grade the answer.
Write me a prompt for summarizing sales calls.
Write a prompt I can reuse in ChatGPT to summarize recorded sales calls from their transcripts.
What I know: our account executives sell payroll software to companies with 50 to 500 staff. After each call, the summary goes into the CRM, where the account executive's manager reads it to decide whether to join the next call. Managers care about budget, the decision-maker, the timeline and any competitor named.
In the prompt, leave clearly labeled blanks such as {{transcript}} for anything that changes from call to call, and do not invent facts about our product or customers. Keep the prompt under 250 words and return it in a code block. A good prompt lets a new account executive get a usable summary on the first try.
Before you write it, ask me up to three questions about anything you would otherwise have to guess.The one-line request returns a long, generic prompt with invented priorities. The second returns a few questions first, perhaps about how long summaries should be, and then a prompt built around the four things managers actually check, with a blank for the transcript.
Make my prompt better. """ [my prompt] """
My current prompt: """ [my prompt] """ The prompt above drafts replies to supplier invoice disputes, and the replies come out too long and sometimes promise credits we have not approved. Rewrite it so a colleague can use it without me. The replies it produces should be under 150 words and promise only what the prompt allows. Keep every fact and rule in my version. Do not add new rules, capital letters for emphasis, or a role line. Return the rewritten prompt in a code block, mark anything you changed with [CHANGED], and explain each change in one line after it. If a rule in my version is ambiguous, list it as a question instead of resolving it yourself.
“Make it better” returns a longer prompt in the model’s own style, and you cannot see what was lost. The second names the two problems to fix, keeps your rules, and marks every change, so you can accept or reject each one.
The line to add
Write a prompt I can reuse in {{assistant}} to {{task}}.
What I know: {{who the output is for, what they do with it, and what matters to them}}.
Leave labeled blanks such as {{input}} for anything that changes each time, and do not invent facts about our work. Keep it under {{number}} words, in plain sentences, with no capital letters for emphasis.
Before you write it, ask me up to three questions about anything you would otherwise have to guess.
Then check the result the way you would check a colleague’s draft. Paste it into the Grader to see which of the ten checks it answers, cross out any “fact” you did not supply, and run it on two or three real inputs in Prompt Tester before you share it.
In ChatGPT, Claude, Gemini and Copilot
Any of the four will write a prompt when asked, and the request above works in each. The vendors build the same idea into their developer tools. OpenAI’s Playground has a Generate button that uses meta-prompts, which incorporate OpenAI’s prompting best practices, to generate or improve prompts (Prompt generation). Anthropic’s guidance suggests asking Claude to evaluate your examples for relevance and diversity, or to generate more from your first set (Prompting best practices).
The same vendors’ guidance is what to check the result against. Anthropic recommends normal language over “CRITICAL: You MUST” (tool usage), and OpenAI recommends removing repeated instructions and examples (Favor leaner prompts). If the prompt is for a different assistant from the one writing it, say which, so the layout suits the reader; the guide to Claude, GPT and Gemini sets the differences side by side.
If you would rather answer questions than review a draft, the Prompt Builder asks you the questions in plain words and lays out the prompt for the model you choose.
Key takeaways
- A model can supply a prompt’s structure; only you can supply its facts.
- Give it the reader, the use and the standard, and ask for questions first.
- Ask for labeled blanks for anything that changes each time.
- Cut invented details, repeated rules and capitals for emphasis.
- Test the prompt on real inputs before anyone relies on it.
Questions
Can ChatGPT write prompts for me?
Yes. It is good at the structure of a prompt: sections, format, rules for missing information. It cannot know your reader, your facts or your standard, so tell it those, ask it to leave labeled blanks for what changes, and ask it to put questions to you before it writes. Then check the result for anything it invented.
What is meta-prompting?
Meta-prompting means using a model to write or improve a prompt instead of asking it for the final answer. OpenAI’s Playground, for example, has a Generate button that uses meta-prompts to draft prompts from a task description. The result is a draft to review and test, not a finished prompt.
Why does the prompt AI wrote for me not work well?
Usually for one of three reasons: it filled gaps with plausible guesses about your work, it grew long with repeated or emphatic rules, or it was never tested on real inputs. Cross out facts you did not supply, cut repetition, and run it on two or three real cases in Prompt Tester.
Prompts to use with this
- Improve a prompt that gives inconsistent resultsYou need a prompt you use often rewritten so it gives reliable, well-formatted answers.