Meta-Prompting 101

A few weeks ago I watched a video that reframed how I think about AI.

Not a product launch. Not a research paper. A creator named Dan Koe walking through how he actually uses Claude and ChatGPT — not as a chat box, but as a system he programs to do exactly what he wants.

The core insight was embarrassingly simple: most people use AI like a slot machine. Type something in, pull the lever, hope for the best. And when the output is mediocre — which it usually is — they blame the AI.

The AI isn’t the problem. The instruction is the problem.

How Most People Use AI (The Slot Machine)

When you type “write me a landing page” into Claude, the AI has to guess almost everything. Your tone. Your audience. Your offer. What makes you different. How your customers talk. Whether you’re casual or corporate.

It guesses by pulling from the most average, most common patterns on the internet. That’s what it was trained on. That’s what it defaults to.

The result sounds like AI. Because it is AI — unguided, averaging everything into nothing.

Here’s the thing: if you hired a human copywriter and gave them the same brief — “write me a landing page” — you’d get the same mediocre result. No context, no voice, no method. Just vibes.

The difference is that a human would push back and ask questions. AI doesn’t push back. It just guesses. And guessing at scale produces slop at scale.

How to Write Expert AI Prompts

To get AI to do something well, you need to teach it exactly how you would do the thing yourself. Not “write me a landing page” but: here’s my offer, here’s the structure I want, here’s an example I love, here’s my customer’s language, here’s what to avoid.

That means writing prompts that are 500 to 2,000 words. Not one sentence. A full set of instructions that removes as many degrees of guessing as possible.

But that creates an obvious problem: what if you don’t know how to do the thing you’re asking AI to do?

This is where Dan Koe’s framework clicked for me. There are four ways to create the expert instructions that make AI actually useful. Click each one to learn more:

The Meta-Prompt: A Prompt That Creates Prompts

You have your expert instructions. But if you just paste them alongside your request — “here’s how to write a good landing page, now write mine” — you’re still missing something: your context.

The AI has the method. It doesn’t have you. It doesn’t know your business, your customers, your voice, your constraints, your story.

The fix is a meta-prompt. A prompt that creates prompts.

Instead of writing the final prompt by hand, you give AI a template for how to build prompts. It asks what you need. It takes your expert instructions. And it produces a prompt that interviews you for your specific context before executing anything.

Meta-Prompting in Practice

Two examples that show how this scales — from a single prompt to an entire system.

Dan Koe’s Method, Summarised

Step 1. Define what you’re building. What does the output look like? Who is it for?

Step 2. Create expert instructions — write your own, ask AI, use a source, or reverse-engineer an example.

Step 3. Feed those instructions into a meta-prompt. Context gathering, execution, refinement — three phases, one prompt.

Step 4. Test. Refine. Every iteration makes the system sharper. The prompt improves every time you use it.

This is the foundation. But a collection of good prompts still needs a home — a place where they work together with persistent context about who you are and what you’re working on.

That’s what Claude Projects are for. And once you have a project, those refined prompts become reusable skills — .md files you attach and reach for whenever you need them.

Next: How to Set Up Claude Projects →

The meta-prompting framework in this article is based on Dan Koe’s video on using AI. I’ve been applying his method to build AI systems for businesses — the guide above is a practical summary. Use it, adapt it, make it yours.