Meta-Prompting: How to Get AI to Write Your Prompts for You

What meta prompting actually is

Meta prompting is asking an AI to write a prompt for you instead of writing it yourself. You describe the job you want done and the shape of the output you need, and the model returns a structured, reusable prompt with a role, constraints, and blank slots for the details that change each time. You then trim it and keep it.

The word “meta” in meta prompting is doing a lot of work and putting people off unnecessarily. Nothing technical is happening. You are briefing a contractor instead of doing the job yourself, and the contractor happens to be very fast and slightly long-winded.

The difference is that the second version describes a machine, not a document. That distinction is the whole of meta prompting, and it is why meta prompting sits a level above ordinary prompt engineering as a skill: you are not learning phrasing tricks, you are commissioning something you’ll use fifty times.

The built-in tools most people miss

Both major AI companies already ship a prompt generator. Anthropic’s sits in the Claude Console alongside a prompt improver, and Anthropic’s own documentation on Console prompting tools describes it as a fix for the blank page problem. OpenAI has an equivalent generator in its developer Playground.

Here is the catch nobody mentions: both live on the developer platform, not in the chat app you actually use. They need a separate account, OpenAI’s asks for card details, and the Claude Console’s template variables don’t work in claude.ai at all. Paying for Claude Pro or ChatGPT Plus does not unlock either one — that’s a subscription for the consumer app, a different product entirely.

So if you are a non-technical professional working in the normal chat window, the built-in generators are not really available to you. The good news is that meta prompting never needed them. Everything below works in an ordinary chat, on the free tier, in any of the four major assistants.

The Commission–Cut–Confirm method, step by step

The Commission–Cut–Confirm approach we teach at PromptPeakAI turns meta prompting into three stages: commission the prompt with a precise brief, cut the draft down hard, then confirm it on one real task before you keep it. Most people do the first stage and skip the other two, which is why they end up with a folder of bloated prompts they never open again.

A meta prompting brief needs the same four things as any good AI input: who the model should be, what it should produce, what it needs to know, and what shape the answer takes. That is the Role, Goal, Context, Format structure applied one level up — you are specifying the prompt, not the document.

The single most useful line you can add is a length cap on the prompt itself. Without it, you are asking a verbose system to be comprehensive, and it will oblige.

Act as a prompt designer. Write a reusable prompt I can paste into a chat AI whenever I need to [DESCRIBE THE RECURRING TASK].

The prompt you write must:
- Assign a clear role suited to [YOUR INDUSTRY OR FUNCTION]
- Ask me for these inputs, in square brackets: [LIST THE 2-4 THINGS THAT CHANGE EACH TIME]
- Specify the output format: [E.G. 200-WORD EMAIL / 5-BULLET SUMMARY / TABLE WITH 3 COLUMNS]
- Contain no more than five rules total
- Be under 200 words

Give me only the prompt itself. No explanation, no preamble, no commentary about your choices.

Those last two lines matter more than they look. “No more than five rules” and “under 200 words” are the constraints that do most of the anti-bloat work before you ever reach the editing stage. If you want the output shape pinned down harder, the same logic used for specifying output format directly applies here.

Read the draft once and delete everything that is not an instruction. Adjectives describing how excellent the AI should be, restatements of the same rule in different words, and any sentence that would not change the output if it vanished — all of it goes.

Fill the brackets with an actual piece of work — a real set of notes, a real email thread — and run it. Invented test inputs are tidier than reality and hide exactly the failure you need to find.

If the output is right, save it. If it is close, fix the one element that caused the gap rather than rewriting the whole thing, which is the same diagnostic habit that works when you need to refine AI output without starting over.

Three meta-prompts for real work tasks

These three are meta prompting briefs, not finished prompts. Paste one in, fill the brackets, and you get a draft prompt back that you then cut. Each takes about a minute to run.

Start with a task you genuinely repeat. Meta prompting a one-off is a waste of a good technique.

Act as a prompt designer for busy managers. Write a reusable prompt that turns a long document into an executive summary for [WHO WILL READ IT — E.G. A REGIONAL DIRECTOR].

The prompt you write must ask me to paste the source document, force a bottom-line-first structure, cap the summary at five bullets, and require that every bullet contains a number or a decision. It must instruct the AI to flag anything the source document does not actually say rather than filling the gap.

Maximum five rules. Under 200 words. Return only the prompt.

Pair that one with the walkthrough on writing executive summaries with ChatGPT, which shows what good output looks like before you automate it.

Act as a prompt designer for an internal communications team. Write a reusable prompt for drafting announcement emails about policy or process changes at [YOUR ORGANISATION TYPE].

The prompt you write must ask me for: the change, the effective date, and what the reader has to do about it. It must enforce a plain reading level, a calm tone, and a maximum of 180 words. It must ban phrases like "we are excited to announce" and any wording that buries the action item below the second paragraph.

Maximum five rules. Under 200 words. Return only the prompt.

Act as a prompt designer. Write a reusable prompt that converts messy meeting notes into a decisions-and-actions list for [YOUR TEAM TYPE].

The prompt you write must ask me to paste the raw notes, then produce two sections only: Decisions Made, and Actions with an owner and a date against each. Anything without a named owner goes into a short Unassigned list rather than being quietly dropped. No summary paragraph, no meeting narrative.

Maximum five rules. Under 200 words. Return only the prompt.

If you want a starting library rather than building from scratch, the free AI work templates cover a good spread of these recurring office tasks already.

Why AI-written prompts come back bloated

Meta prompting produces bloated drafts because language models associate the request for a “good prompt” with length and formality, so they pad with descriptive adjectives, restate rules, and stack constraints that were never asked for. The result reads impressively and performs worse than a lean version of the same instructions.

This is not a fringe complaint. Anthropic’s documentation says plainly that its own prompt improver produces templates that give longer, more thorough, but slower responses, and recommends simpler prompts where speed or cost matters. When the company that built the tool tells you its output runs long, believe it.

Meta prompting has a second, sharper failure mode. Ask for a comprehensive prompt and the model will sometimes invent constraints that sound sensible but describe features or steps that do not exist. That is ordinary AI hallucination pointed at your instructions rather than your content, and it is harder to spot because a fake rule looks exactly like a real one.

The pruning pass, rule by rule

The pruning pass is the stage that makes meta prompting pay. Five minutes with these four rules turns a draft into something worth saving. Work top to bottom.

• Halve it, or explain why not. If you cannot cut a 600-word draft to around 300, you have probably kept padding you mistook for nuance. The target is a rough halving, not a precise ratio.
• Delete every adjective that does not change behaviour. “World-class”, “highly experienced”, “meticulous” — none of these alter the output. “Under 180 words” does.
• Cap it at five rules. Beyond roughly five constraints, models start quietly dropping some. Five you can trust beats twelve you cannot.
• Check every rule is real. If the prompt references a button, menu, or feature, confirm it exists before you save it.

You can also hand the pruning back to the AI, which works surprisingly well because cutting is an easier judgement than writing.

Below is a prompt. Cut it by roughly half without losing a single instruction that changes the output.

Delete: adjectives about how skilled or thorough you should be, restated rules, and any sentence that would not change the result if removed. Keep: the role, the required inputs, the output format, and hard limits.

If the prompt contains more than five rules, tell me which ones you would drop and why. Then give me the cut version only.

PROMPT TO CUT:
[PASTE YOUR DRAFT HERE]

Then read the result yourself. It is a decent editor and a poor judge of what your job actually requires, so the final call on which five rules survive stays with you.

Save the survivors somewhere you’ll find them, because meta prompting only pays back on the second use. A personal AI workflow library is the low-effort version; for prompts you run constantly, promoting them into ChatGPT custom instructions or the equivalent project settings removes the paste step entirely.

When meta prompting is the wrong move

Meta prompting earns its keep on tasks you repeat with a fixed output shape. For a one-off question, it is slower than just asking — you spend three minutes commissioning a tool you will never open again.

One model-specific wrinkle changes how hard you should push meta prompting. OpenAI’s prompt engineering guide notes that reasoning models tend to do better with high-level guidance, while standard models benefit from very precise instructions. Piling twelve constraints onto a reasoning model can actively make things worse — another reason the five-rule cap is not arbitrary.

What meta prompting cannot fix

Worth naming the quiet worry too: no, nobody can tell you used meta prompting to write your prompt. The prompt never leaves your chat window. What people notice is output that reads like a template, and the pruning pass is precisely what prevents that.

One real caution applies to meta prompting itself. Describing your recurring task often means describing how your organisation works, and that detail goes into a consumer chat product.

When a generated prompt still fails

Most meta prompting failures come back to the same handful of causes, and each has a one-line fix rather than a rewrite.

If the prompt is broadly right but the tone is off, the fastest repair is showing rather than telling. Paste two examples of output you liked into the prompt — the few-shot prompting technique does more for voice matching than any adjective you could add.

The habit worth building is not writing better prompts. Meta prompting works because it treats the AI’s first draft as a starting point in every direction — including when the thing it drafted was the instruction itself.