The order matters, and so does the pacing. Each answer becomes context for the next prompt, so Claude is critiquing progressively more of its own thinking as you go. Paste them all at once and you get one blended, watered-down response instead of four distinct passes.
A first answer arrives with no signal about which parts are solid and which parts got filled in. This prompt forces the answer to be broken into separate claims with a number attached to each, so you can see the soft bits.
It works better than it sounds like it should. Stanford researchers found that when models fine-tuned with human feedback (ChatGPT, GPT-4 and Claude among them) are asked to say their confidence out loud, those spoken confidence scores are "typically better-calibrated than the model's conditional probabilities", often cutting calibration error by around half (Tian et al., Just Ask for Calibration). Asking is genuinely informative, as long as you ask per claim rather than for the answer as a whole.
Before I act on this, rate your confidence in the answer you just gave me out of 10, and explain the number.
Break your answer into its separate claims and give each claim its own score, because an answer is almost always a mix of things you are certain about and things you filled in to sound complete. For anything you score 7 or below, say plainly what you are unsure about and why it is shaky. Then tell me exactly what would raise each score: which detail from me you are missing, or which source you would need to check. Do not rewrite the answer yet, I only want the honest read on it.The first answer is built for the situation Claude assumed you were in. This prompt hunts for the version of your situation where that answer quietly becomes the wrong advice, which is the failure that actually costs you something.
Now argue against your own answer. What could fall apart here, and what is the version of this answer that turns out to be wrong?
Give me specific failure cases rather than general caution: the assumptions you made about my situation that might not hold, the edge cases you skipped over, and the exact conditions under which your advice becomes the wrong advice. For each one, tell me how I would notice it happening early, while it is still cheap to fix. Then name the single most likely failure of the set, and tell me what you would change in the answer to protect against that one.A general answer is built from what is written down most often, which is rarely what a practitioner would actually do. Naming a specific role, with a specific amount of hands-on experience, pulls a different set of considerations into the answer: the standard practice, the thing that always goes wrong in real cases, the shortcut nobody bothers to write down.
Answer this again from scratch as a [role, for example: senior tax accountant who has filed hundreds of these / performance marketer who has spent seven figures on Meta / employment lawyer who negotiates these contracts weekly], and start by telling me what your first answer missed.
Do not politely restate the earlier answer, redo the thinking with that expertise in front of it. Name the things a practitioner would weigh that a general answer would not: the standard practice in this field, the thing that always goes wrong in real cases, and the shortcut people who do this every week actually use. Then give me a short list of what changed between your two answers, and say which version you would stake your professional reputation on and why.This is the one that earns the whole routine, and it's the one people skip. The first three prompts are Claude checking Claude, and there's a real ceiling on that: a widely cited ICLR 2024 paper found that "LLMs struggle to self-correct their responses without external feedback, and at times, their performance even degrades after self-correction" (Huang et al., Large Language Models Cannot Self-Correct Reasoning Yet). So the fourth prompt stops asking the model to grade itself and asks it to hand you the external check instead.
Now hand me a way to check this without you. How would I verify each important claim in your answer myself?
For every factual claim, give me the primary source I should look at (the specific document, the official body, the page or section, or the exact search phrase that finds it), plus what that source needs to say for your claim to hold up. Where something cannot be sourced, label it clearly as your judgement rather than fact instead of leaving me to guess which is which. Finish with the three fastest checks that would catch the biggest error if you are wrong, ordered by what I should do first.They are not four ways of asking the same thing, which is why stacking them works and why dropping one leaves a hole.
Running four extra prompts on every message would be exhausting and mostly wasted. The test is for answers you're about to act on.
Once you've run this a few times, stop pasting it. A skill is a small folder with a SKILL.md inside that holds a name, a description of when to use it, and your instructions. Claude reads the description automatically and pulls in the full skill when your request matches, so the routine becomes one word. Skills work on every plan, Free through Enterprise, and you don't write any code.
Build me a reusable Claude skill called "pressure" that runs a four-pass pressure test on an answer you have just given me.
Use skill-creator to write it. The SKILL.md description should say to use it whenever I type /pressure or ask you to pressure test, sanity check or stress test an answer. The instructions should run four passes in order and label each one clearly in your reply: pass one scores your confidence per individual claim out of 10 and names what would raise each score, pass two argues against your own answer with specific failure cases and early warning signs, pass three redoes the answer from scratch as a named expert practitioner and lists what changed, and pass four gives me primary sources I can check myself plus the three fastest checks that would catch the biggest error. Finish every run with a short "what I would change" summary and a plain flag on anything I should not act on yet. Then package it as a zipped folder I can upload.~/.claude/skills/ instead. Anthropic's create custom skills article has the same steps with screenshots./pressure after any answer that matters. Four passes, one word, no pasting.
Just Ask for Calibration (Tian et al.), on verbalized confidence beating raw model probabilities
Large Language Models Cannot Self-Correct Reasoning Yet (Huang et al., ICLR 2024)
Stanford HAI, AI on Trial: hallucination rates in purpose-built legal research tools
Vectara's hallucination leaderboard, tracking rates across current models