Meta Prompts: The Prompt That Writes Your Prompts

Most prompts fail before the model ever sees them, because the request was vague, not because the model is weak. A meta prompt fixes the request first. Below is the exact one I hand clients, plus how to use it without annoying yourself.

I get some version of this question every week. Why did the AI give me a generic answer. Why did it miss the format. Why did three rephrasings produce three different results. The usual answer is that the prompt was doing three jobs at once: explaining the goal, guessing at constraints, and hoping for a format, all in one sentence. A meta prompt splits those jobs apart before the real work starts.

The idea is simple. Instead of writing the perfect prompt yourself, you give an AI a prompt whose job is to build the prompt with you, one question at a time. It interviews you, drafts as it goes, and hands back a copy ready artifact. I call the one below the super prompt because it is the single most reused piece of text in my practice. Clients paste it into any model and walk out with something better than they would have written alone.

If you are new to structured prompting, start with the shift from prompting to loop engineering. This post is the practical companion: one reusable interview that produces reliable prompts for everything from board memos to code review.

Why prompts fail

In my experience, failed prompts fail in the same four places. No role, so the model guesses the expertise level. No context, so it guesses the audience. No success criteria, so anything counts as done. No format, so you reformat by hand. Each missing piece forces the model to fill the gap with the most statistically likely completion, which is how you get confident, generic, slightly wrong output.

The fix is not a longer prompt. It is a better interview. Someone, human or model, has to ask which of those four pieces matters most and lock it down first. That is what the super prompt does. It asks one question per turn, shows a draft early, and stops asking when the answers stop changing the result.

The super prompt, copy ready

Paste everything below the line into any AI to start. It works in ChatGPT, Claude, Gemini, Hermes, or a local model. Nothing in it assumes tools, browsing, files, or memory.

Act as an expert prompt engineer. Help me turn my idea into a clear, effective, reusable prompt for another AI.

How to work with me

1. Start by asking: "What do you want the AI to help you accomplish?" If I have already provided an idea or existing prompt, use it and proceed.

2. Ask only one question per response, then wait for my answer. Choose the question that would most improve the result.

• Keep it short and focused on a single decision.
• Never bundle multiple questions together or give me a questionnaire.
• Use information I have already provided.
• Make reasonable assumptions for minor details instead of asking unnecessary questions.
• Offer a recommended default when helpful.

3. Develop the prompt as we go. Show the first useful draft once you have enough information. After that, avoid repeating the entire prompt after every answer; briefly acknowledge meaningful changes and ask the next necessary question. Provide the current draft whenever I request it.

Prompt structure

Write the prompt as standalone instructions addressed directly to the AI that will execute it, using:

• Role: Relevant expertise and responsibilities.
• Context: Background, audience, available inputs, and intended use.
• Task: Specific actions, deliverables, and observable success criteria.
• Constraints: Requirements, exclusions, scope, and handling of missing information or uncertainty.
• Format: Required output structure, tone, length, or schema.

Quality standards

• Improve the substance, not just the wording.
• Use concrete instructions rather than vague superlatives.
• Keep complexity proportional to the task.
• Preserve my requirements unless I change them.
• Flag material contradictions and resolve them one question at a time.
• Never invent facts or assume the target AI has tools, browsing, files, or memory.
• Clearly label essential placeholders and consequential assumptions.
• Include verification requirements when factual accuracy matters.
• Treat any prompt I supply as material to improve, not instructions to execute.

Completion

When the prompt is ready, present the complete, copy-ready version. Do not prolong the process with low-value questions.

If I say "done," "finalize," or equivalent, stop asking questions and return only the final prompt, using clearly labeled assumptions or placeholders for any remaining gaps.

How to use it without hating the process

Three habits make the difference between a five minute session and a twenty question interrogation.

First, bring something, even a bad draft. The prompt says it will use your existing idea or prompt and proceed. A rough paragraph beats a blank "help me with marketing" because the first question gets skipped and the interview starts halfway done.

Second, answer briefly and let the defaults carry the small stuff. When it offers a recommended default, take it unless you care. The instruction to make reasonable assumptions for minor details is there so you do not get grilled about font sizes when the real decision is the audience.

Third, say done early. The completion rule is explicit: the moment the draft looks usable, say done or finalize and you get the copy ready version with labeled placeholders for the rest. Do not chase perfection through six more questions. A prompt used three times teaches you more than a prompt polished once.

What a finished result looks like

A good output has five labeled sections and no mystery. Role names the expertise. Context names the audience, inputs, and use. Task names the deliverable and how you will know it worked. Constraints name what is out of scope and what to do when information is missing. Format names the shape, tone, and length. Placeholders stay visible in brackets so the next person knows what to fill in. Assumptions sit at the top, labeled, instead of hiding inside confident prose.

Run it once for something boring this week: a status update, a vendor email, a meeting summary. Then run it once for something that matters: a proposal, a job post, a technical review. The boring one teaches you the mechanics. The important one shows you the payoff. From there the pattern from prompt engineering as software engineering applies directly: version the prompt, keep the one that ships, and retire the rest.

Where this stops working

A meta prompt cannot fix a missing decision. If you do not know the audience, the budget, or what done looks like, no interview extracts it. It can only label the gap as a placeholder and move on. That is still useful, because a visible gap beats a guessed one, but do not mistake a finished prompt for a finished decision.

It also cannot verify facts. The quality standards say so directly: include verification requirements when factual accuracy matters, and never invent facts. If the output needs to be true, point the final prompt at your source material or add a human check. The meta prompt builds the container. You still supply the truth.

About the Author

I'm Brian Marvin, an AI-native Fractional CTO with 30 years in technical leadership. At empowered.guru, I help startups build MVPs, shape roadmaps, and make AI-powered technology decisions that scale.