When to use it: Before you paste an AI answer into anything that matters: a deck, a client email, a report.
When to avoid it: Low-stakes internal scratch work. Running it on everything turns into busywork. Save it for outputs headed somewhere real.
The prompt:
Pressure-test your last answer before I use it. Do all four, briefly:
1. Confidence: rate how sure you are, 1-10, and say what would lower it.
2. Weak points: flag every specific number, date, name, quote, or source
you gave. For each, say "verified from my knowledge" or "you should
check this."
3. Assumption check: what did you assume about my situation that, if wrong,
breaks the answer?
4. Verify steps: tell me the fastest way to confirm the flagged items,
what to search, what to look at.
Don't defend the answer. Just show me where it's soft.
Exact steps:
- After the AI gives you an answer, paste the pressure test into the same chat.
- Read what it flags.
- Check only the flagged items against a real source. Send.
Expected result: The model annotates its own sureness. The soft spots (numbers, dates, names, sources) become visible in about 10 seconds, so you check those instead of everything.
Where it breaks: It’s checking its own work, so it can be confidently wrong about being wrong. A “verified” is a hint to relax, not a guarantee. The confidence score (“8/10”) is a way to rank what to check first, not a real probability. Skip the “only answer if confidence above 0.8” trick going around; it just makes the model write a number above 0.8 and keep going.
What to check before using it: Anything headed to a client, and any number that matters, still gets a real-world check, label or no label.