Interactive eBook BuilderInteractive eBook Builder
$59 · unlimited access, unlimited eBooksGet it for $59

· By · 10 min read

Meta-Prompting: Use AI to Write Better Prompts

promptsaiwritingebook

Meta-prompting is the practice of using a language model to generate or improve a prompt, then using that improved prompt on the actual job. OpenAI’s cookbook article Meta Prompting: A Guide to Automated Prompt Optimization describes it as a way to take a basic prompt and refine it so later outputs are more specific and complete. The page is now marked as archived and may reference older models, so treat it as a method, not as a current model ranking. The useful idea has not changed: one request works on the brief, a later request works on the book.

That is different from turning a messy brain dump into a prompt, which is something you can do by hand, and different from improving a prompt when the output is generic, which is a human loop: build a small test set, change one variable, and score the result. Meta-prompting can feed that loop. It should not replace it. A rewritten prompt can sound more professional and still be vaguer, longer, or untrue to your notes.

For eBook and workbook projects, Interactive eBook Builder is a $59 one-time Notion AI template with Spec Builder, Drafter, Polisher and SEO BOT already set up. The Prompt Engineering Lab in the Writers Workshop is the right place to test a draft brief and a rewritten brief on the same chapter job. Notion AI access and usage are separate from the purchase. The examples below are teaching prompts written for this article. They are not paid Builder prompts, and they are not recorded model outputs.

What should a meta-prompt actually ask for?

A weak meta-prompt says “make this better” and accepts the first rewrite. A better one tells the model what “better” means, asks it to find gaps, and keeps you in the loop.

Ask the model to do four things, in order:

  1. Restate the job in one sentence, as it currently understands it.
  2. List missing information: audience, outcome, constraints, sources, format, length and tone.
  3. Ask you clarifying questions for anything it would otherwise have to invent.
  4. Only then, draft a revised prompt you can edit, with clearly marked sections.

If you skip step 3, the model will often fill gaps with generic defaults: a broad audience, a “comprehensive” chapter, or a format you did not want.

OpenAI’s archived cookbook starts from a simple summarization prompt and uses a stronger model to add structure and evaluation criteria. You do not need two different products to use the method. You do need two passes: one that works on the prompt, and one that runs it.

When does meta-prompting help, and when does it get in the way?

It helps when you have a rough brief and you can tell, by looking, that it is missing pieces: no reader, no length, no source limit, no definition of done. It also helps when a prompt has grown into a long speech and you want a cleaner structure with headings or tags.

It gets in the way when you let the rewrite add goals you never had, or when you treat the new prompt as finished because it is longer. It also gets in the way if you change the prompt and the model and the examples all at once, because you will not know what helped.

The rewritten prompt is a draft. You still have to read it, cut anything you cannot support, and test it.

Example 1: turn a vague chapter brief into a testable prompt

Here is a typical starting prompt:

Write a chapter about pricing for designers.

A useful meta-prompt does not write the chapter.

Here is a draft prompt I want to improve. Do not write the chapter.

Draft prompt:
"Write a chapter about pricing for designers."

1. Restate the job you think this prompt is asking for.
2. List the missing pieces (reader, outcome, constraints, sources,
   format, length, tone).
3. Ask me up to six clarifying questions. If a question would change
   the book, say so.
4. Stop. Wait for my answers before you rewrite the prompt.

After you answer, continue:

Here are my answers:
- Reader: new freelance designers who undercharge.
- Outcome: name a floor rate before the next inquiry, not a full menu.
- Sources: only the notes I will paste; no invented averages.
- Length: 700 to 900 words, ending with a fill-in line.
- Figure I can use: a $90 floor for a one-page logo revision.

Now rewrite the prompt. Use labeled sections for Goal, Reader, Notes,
Constraints, and Success. Keep it brief. Do not add research tasks,
marketing advice, or a package-menu section. Return only the revised
prompt, plus a three-line list of what you still do not know.

Read that rewrite before you run it. If it has quietly added “include industry benchmarks,” cut that line.

Example 2: use a meta-prompt to tighten a bloated brief

Prompts rot by accumulation. A month later, the same chapter brief may include leftover instructions for a different book.

Here is a prompt I have been reusing. It has gotten long, and I think
some lines are leftovers. Do not run the task.

[paste the long prompt]

Please:
- Group the instructions into Goal, Reader, Standing rules, Notes,
  Output format, and Success.
- Flag any instruction that conflicts with another.
- Flag any instruction that looks generic or copied from a different
  job (for example, a children's workbook rule inside an adult pricing
  book).
- Propose a shorter prompt that keeps only the rules I need for this
  chapter.

Do not add new rules. If you are unsure whether a line matters, list
it under "Decide" instead of deleting it quietly.

This is meta-prompting as editing, not as invention. You stay responsible for the cuts.

Example 3: generate a prompt, then test it against a failure case

A rewrite is not finished until it fails safely. After you accept a revised brief, run it twice: once on a normal chapter, and once on a brief that should force a “need from me” response, such as missing figures or conflicting prices.

I will test two prompts on the same failure case.

Failure case notes:
- I mentioned a $75 floor in one sentence and a $120 floor in another.
- I have no survey data.
- I have not decided which figure to publish.

The prompt under test is:
[paste prompt A or prompt B]

Run only this prompt on the failure case. I want to see whether it
flags the conflict or invents a compromise. Then stop.

Score the two outputs with a small rubric you wrote in advance, such as the one in the generic-answers article: does it serve the stated reader, stay inside the notes, flag gaps, and match the requested format? If the “improved” prompt papers over the conflict more smoothly, it is not better. It is worse.

What should you refuse to let the rewrite do?

A common failure is a prompt that comes back longer, more formal, and less true to the book. Watch for these additions and cut them on sight:

If you are rewriting a brief for a reasoning model, say that in the meta-prompt. Ask for a short goal, constraints and success criteria, not a tutorial on how to think. Then test the result the same way you would test any other change: one variable, a normal case and a failure case.

How should you use this inside a writing workspace?

Keep the original prompt, the questions, your answers, the rewrite and the test notes together. Spec Builder can hold the accepted brief for a chapter. Drafter should run that brief, not a fresh rewrite of its own on every pass, unless you are deliberately testing a change. Polisher needs a narrower prompt, so do not let a meta-prompt for drafting bleed into editing.

The Prompt Engineering Lab is built for the comparison: one instruction changes, you predict the effect, and you look at a normal case and a failure case. Use the Lab, or the same habit by hand, so a polished rewrite does not skip the test. If you are learning prompt engineering by building, this is a good second exercise after a single checklist prompt: improve the brief, then prove that the new brief behaves better.

FAQ

What is meta-prompting?

Meta-prompting means using a model to write or revise a prompt, then using the resulting brief on the actual task. OpenAI’s cookbook describes it as a way to refine a basic prompt, often with a more capable model helping to rewrite a brief for another model or a later run. In everyday chat, it can be as simple as: “Critique this prompt, ask me questions, then rewrite it.” The rewrite is only a candidate until you test it.

Can ChatGPT improve my prompt?

It can point out missing audience, constraints, sources and format, and it can propose a clearer structure. It cannot tell, by itself, whether the new brief will stay faithful to your notes. Answer the clarifying questions yourself, edit the rewrite, and run both the old and new prompts on the same examples. If you want a systematic loop, use a small test set and a rubric rather than a single impression.

Should I let the model rewrite the prompt and run it in one step?

No. If the same reply rewrites the prompt and writes the chapter, you will not have approved the brief, and you will not know whether a bad chapter came from a bad rewrite. Split the work: improve the prompt, edit it, then run it. That also makes it easier to save the brief and reuse it on the next chapter.

Before you hit send

A better-sounding prompt is a draft. A better-tested prompt is the one you should reuse.

Ship niche eBooks with agents already built.

Get the Interactive eBook Builder for $59