How to Ask AI to Write the Prompt for You
Let the AI interview you, draft the prompt, then break it on purpose. A three-step method with a reusable template for prompts you will run again and again.
How to ask AI to write the prompt for you comes down to not saying write me a prompt for X. Tell it your goal, then ask it to interview you with questions until it has enough to draft the prompt, and finally test the draft on one awkward input. The interview is where the quality comes from, because the model tends to fill gaps with generic assumptions if you let it.
This is sometimes called meta-prompting: using a model to produce the instructions for a model. It works well for prompts you will reuse, like a weekly report or a support reply, and badly when used as a one-click shortcut. Here is the method, a template, and the test that tells you whether the result is worth keeping.
Why one-line requests give weak prompts
Ask for a prompt that summarizes meeting notes and you will get a tidy, generic template: summarize the notes, list action items, keep it concise. It is fine and it is the same prompt everyone gets. It does not know who reads the summary, what a good one looks like for you, or what usually goes wrong.
Anthropic's prompting guide offers a handy sanity check for any prompt: show it to a colleague with minimal context and ask them to follow it. If they would be confused, the model will be too. The interview step is designed to put the missing context on the page.
Step 1: make the AI interview you
Paste this into a fresh chat, with your goal in the brackets:
I want a reusable prompt for this job: [describe the task in one or two sentences].
Do not write the prompt yet. First ask me up to 8 questions, one at a time,
about things that would change how the prompt is written. Cover: who reads the
output, what a good result looks like, what a bad result looks like, the input
format, any limits on length or tone, and anything the AI must never do.
After my answers, ask whether I want to add anything. Then write the prompt.Answer in specifics. Instead of a clear tone, say plain, no jargon, written for a shop owner who skims on a phone. Paste a real example of a good past output if you have one. A single example carries more information than a paragraph of adjectives.
Step 2: get the draft, then read it like a stranger
When the draft arrives, check it against this list before you run it.
Does it state the reader and the goal in the first lines?
Does it describe the input and say where it will be pasted?
Does it say what to do when information is missing, such as asking or marking it unknown?
Does it specify the output shape, such as headings, a table or a word limit?
Does it include at least one of your own examples or constraints, not only generic advice?
If two or more answers are no, send the draft back with the gaps named. Do not rewrite it by hand yet. Asking the model to fix specific gaps is faster than editing from scratch, and you learn which of your answers were too thin.
Step 3: break it on purpose
A prompt that works on the easy input proves little. Pick the nastiest realistic input you have: the messy notes, the angry customer, the invoice with a missing field. Run the draft on it, in a fresh chat, three times.
What you see | What it means | What to do |
|---|---|---|
Good output all three times | The prompt is sturdy for this case | Save it and move to a second awkward input |
Good once, bad twice | The prompt leaves a decision open | Add a rule for that decision, then re-run |
Confident but wrong | A constraint is missing | Add the constraint and a do-not line for this mistake |
Ignores a rule | The rule is buried or contradicted | Move it earlier, state it once, remove conflicts |
Feed the failure back into the same chat that wrote the draft: here is the input, here is the bad output, update the prompt so this does not happen. Then re-run. Two rounds are usually enough for a prompt you will use many times.
A reusable template for the final prompt
Role and reader: [who the output is for, one line]
Goal: [what the output must achieve]
Input: [what will be pasted below, and its format]
Rules:
- [rule 1, written as an action]
- [rule 2]
If something is missing: [ask / mark unknown / skip]
Output format: [headings, table, word limit]
Example of a good output:
[paste one]If your prompt is longer than a page, separate the parts with labeled sections. The approach is explained in XML tags in AI prompts, and once a prompt is working, how to version your prompts stops later edits from quietly breaking it.
When not to use AI to write your prompt
Skip it for a one-off question. The interview costs more time than the answer saves. Skip it, too, when you cannot judge the result, since a polished but wrong prompt is harder to spot than a rough one.
The broader toolkit is in the prompt engineering guide. For prompts tied to code builders, see how to write prompts for AI app builders, and for the reasoning behind giving examples, prompt AI with a worked example.
FAQ
What is meta-prompting?
Using an AI model to write or improve the prompt you will give to an AI model. It can be as simple as asking for a draft, or as structured as the interview method above.
Can ChatGPT write better prompts than I can?
Only with your input. It does not know your reader, your standards or your past failures until you tell it, which is why the interview step comes first.
How do I know the AI-written prompt is good?
Run it on your hardest realistic input several times in fresh chats. A good prompt gives consistently acceptable output on the awkward case, not just the easy one.
Should I keep the prompt the AI wrote as is?
Read it first, cut anything generic, and keep your own examples and rules. The parts that came from your answers are the valuable ones.
How did this land?
About the author
Developer Advocate
Steve builds something with Swarmz every week and writes up what worked, what broke, and what he'd do differently. Tutorials and hands-on guides are his lane.
