菜单
Use when testing a hypothesis, debugging, validating a fix, designing experiments or test suites, or when checks keep passing. Applies Popper's falsification and Wason's 2-4-6 task — a test that can't fail isn't a test, and confirming instances prove little.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when tempted to say "I'm being objective", when attributing others' views to bias while trusting your own introspection, or after learning about a bias and concluding it doesn't apply to you. Applies Pronin, Lin & Ross (2002) — the bias blind spot and the introspection illusion.
Use when stating confidence ("I'm sure", "definitely", "99%"), giving confidence intervals, making probabilistic claims, or reviewing a track record of predictions. Applies overconfidence calibration research (Lichtenstein & Fischhoff; Alpert & Raiffa) — subjective certainty systematically exceeds accuracy.
Use before deleting code, removing a check, simplifying a process, dropping a config flag, or removing anything that "serves no purpose" — especially in unfamiliar or legacy systems. Applies Chesterton's Fence (G.K. Chesterton, 1929): don't remove it until you know why it's there.
Use when evidence seems to confirm an existing belief, when research keeps agreeing with the hypothesis, when interpreting ambiguous data, or when a first explanation feels satisfying. Applies Lord, Lepper & Preston (1984) — the one debiasing instruction shown to work: consider the opposite.
Use when writing docs, error messages, onboarding, tutorials, API designs, UX copy, or teaching — or when something you built seems self-explanatory. Applies the curse of knowledge (Camerer, Loewenstein & Weber 1989; Newton's tappers-and-listeners 1990): expertise makes you unable to simulate not-knowing.
Use when operating in an unfamiliar domain, when something "looks right" but hasn't been verified, when confidence is high without external checks, or when evaluating one's own competence at a task. Applies Kruger & Dunning (1999) — the skills needed to perform are the skills needed to judge performance.
| name | blind-falsification |
| description | Use when testing a hypothesis, debugging, validating a fix, designing experiments or test suites, or when checks keep passing. Applies Popper's falsification and Wason's 2-4-6 task — a test that can't fail isn't a test, and confirming instances prove little. |
Karl Popper (Logik der Forschung, 1934): a claim earns credibility only from surviving attempts to refute it — confirming instances are cheap and unbounded. Wason's 2-4-6 experiment (1960) showed the default human strategy: given a rule to discover, people propose only sequences their hypothesis says are valid, receive "yes" after "yes", and announce the wrong rule with confidence. They never propose a sequence their hypothesis says should FAIL.
The blind spot: a green test you designed to pass tells you almost nothing. The information is in the test designed to fail.
| Situation | Falsification move |
|---|---|
| "Fix verified — it works now" | Revert fix → test must fail → unrevert |
| Hypothesis about the root cause | Predict something it forbids; check that |
| All tests green | Add the test that SHOULD fail (wrong input, boundary) |
| "Can't reproduce" | Define what observation would count as reproduced |
