What is meta-prompting, and why does it improve AI output?

If I want a deep research report, I don't ask AI for the report. I ask it for the prompt that gets the report. That one habit is the core framework I teach non-technical people.

"Don't prompt AI. If you want something serious, you don't ask AI for it, you ask AI for the prompt to get it."

What is meta-prompting, exactly?

It's a two-step. Step one, you tell the AI what you want and ask it to write the best possible prompt for that job. You tell it explicitly not to do the task yet. Step two, you take the prompt it wrote and run it, usually in a fresh chat or in a different mode like deep research.

The AI is a better prompt writer than you are. It knows what context, format and constraints it needs. You just have to let it ask for them.

Why does it work better than asking directly?

Most people still use AI like Google. In Google, typing too much gets you worse results. In any large language model it's the opposite: the more information you give it, the better.

When I ramble a request out loud, I give the model maybe three messy sentences of real intent. When it turns that into a prompt, it adds structure: the scope, the angles to cover, the output format, what to leave out. That's way better context than the three sentences I said. Often I don't even read the prompt it writes. I just paste it in.

What does it look like in practice?

Here's a real one I demoed for a room of insurance agents. I said something like: write me a prompt, I don't want you to do anything, I just want the prompt for deep research on agent compensation in all respects, what changed in 2026, and a projection from 2027 to 2030.

Then I pasted that prompt into ChatGPT's deep research mode. Deep research takes about ten minutes and reads 100-plus websites. Big consultancies used to charge around $20,000 for reports like that. Now you get one in about 15 minutes.

Another one, for market research: I talked to regular ChatGPT, described what I wanted, and asked it to write the deep research prompt. The ask was to find every newsletter, media company or YouTube channel specifically helping restaurant owners run their businesses better, and bring back a list with follower and subscriber counts. You don't start with deep research. You start deep research once you have the prompt.

It works for building an app, too. This is how I tell non-developers to start. Open ChatGPT and say: I need your help building an app. I'm about to give you all the information about it. Your job is to write me a detailed prompt for another model so it can one-shot build this app. Then hit the mic and brain dump everything you think about the app.

You're not asking for the app. You're asking for the prompt. I take that prompt and paste it into a builder like Google AI Studio or v0.

How do I make the prompt better before I run it?

Don't take the first version. Push back once:

This looks pretty good, but I think you can do way better.
Review your work, score it, find everything you may have gotten wrong,
and then redo it the better way.

It takes about an extra minute, and it comes back with a better version. Ask any model whether that was the best it could do and whether it got anything wrong, and it'll say yes.

The same trick works on any answer, not just prompts. My simplest quality habit is to stay in the same chat and ask:

Is this really the best you can do?
Please give yourself a review and tell me what you missed.

Every single time it says something like: this is pretty good, but I missed this, this and this. Then I say go do those. Sometimes I've done it three times. Even once gets you a much better response. Matt Diehl made the point on my podcast that a lot of "art of prompting" advice boils down to this same self-review step.

"I would recommend always just do that at minimum in any LLM you're in. Just say, tell me what you missed and then fix it."

For something that really matters, take it a step further and have a different model critique it. I once had a plan reviewed by a tool set up as a product-requirements specialist. It told me the plan missed the user journey and some UI and UX complications. I told it to fix those things, and it did.

What prompt do I copy to start?

The shape of step one, ready to paste:

Here's what I'm trying to do: [brain dump everything, messy is fine].
Don't do the task yet. Write me the best possible prompt for it.
Include the context, the exact output I want back, and what to avoid.
Ask me anything you need to know first.

For research, swap the second line for: "I don't want you to do anything. I just want a deep research prompt on [topic], covering [angles], with [time frame]."

How do I get the most out of it?

1. Talk, don't type, for step one. I use Wispr Flow for voice dictation. Talking strips out the self-censoring you do when you type, so the AI gets more raw material to build the prompt from. More on that in should you talk or type to AI.
2. Run step two somewhere clean. A fresh chat, a deep research run, or a coding agent. Don't bury a great prompt at the bottom of a long, drifting conversation.
3. Use it for planning builds, too. Same idea, bigger job: plan with one AI, build with another.
4. Save what works. A guest on my podcast recommends saving a solved workflow as a reusable skill so he never has to solve it twice. Good prompts deserve the same treatment.

When should I skip it?

For quick, throwaway questions, just ask. Meta-prompting is for anything serious: research you'll act on, a build, a document other people will read. If you'd be annoyed redoing it, write the prompt first.