Most AI outputs fail not because the model lacks knowledge, but because it jumps straight to answering. It pattern-matches to a likely response and runs with it — skipping the part where it actually reasons through your specific situation. The meta-cognitive reasoning prompt fixes this by forcing the model to surface its own assumptions, identify what it might be getting wrong, and pressure-test its logic before it gives you an answer.
The mechanism is simple: you instruct the model to think about how it is thinking, not just what it thinks. This slows the response down in a useful way. You get fewer confident-sounding wrong answers and more honest, structured reasoning. It works across every major AI tool — ChatGPT, Claude, Gemini, Perplexity — because the underlying principle is universal to large language models.
The Prompt
Before you answer, I want you to do the following:
1. Identify the core question I'm actually asking (in one sentence).
2. List the key assumptions you're making to answer it.
3. Flag any assumption that might be wrong, incomplete, or context-dependent — and explain why.
4. Note any information you would need that I haven't given you, and how the answer might change if that information were different.
5. Then give your answer — clearly labeled — based on your best reasoning after that review.
Do not skip any of these steps. Do not collapse them into a quick paragraph. Work through each one visibly.
Here is my question: [PASTE YOUR QUESTION HERE]When to use it
This prompt is not for simple lookups. Use it when the stakes are high enough that a fast but slightly wrong answer costs you time, money, or credibility. Here are the situations where it earns its keep:
- Pricing or positioning decisions — you want the model to surface what it doesn't know about your market before it tells you what to charge or how to position.
- Reviewing a strategy or business plan — instead of getting a polished endorsement, you get a genuine stress-test of the logic.
- Technical implementation choices — the model flags where its answer depends on your specific stack, team size, or constraints rather than assuming a generic setup.
- Drafting copy or a pitch that will be scrutinized — the model catches its own weak assumptions about your audience before you send it to a real person.
- Any time a previous AI answer felt too neat or confident — run it through this prompt to expose what got glossed over the first time.