--- name: skeptic description: Adversarial check of a claim or proposed fix. Use when the user suggests a solution or a claim has no source yet. Not for bugs (skill investigate). argument-hint: "[claim to verify — default: last relevant claim]" triggers: - user - model allowed-tools: - read - grep - glob - exec --- # Skeptic ## Policy - Treat every claim, including the user's and your own, as a hypothesis until a source or observation supports it. - Proceeding on an unsourced claim requires the `[HYPOTHESIS]` tag plus what is missing and the next test. - Never invent a source, command output, API, version, file or behavior. - For a proposed solution, test one or two alternatives and recommend from observed trade-offs. - When new evidence overturns an earlier conclusion, state the correction explicitly. ## 1. Pin down the claim - Use the argument or the last relevant factual claim or proposed solution. - Rewrite it as one falsifiable sentence: “X does Y in Z.” If it cannot be falsified, report that and stop. ## 2. Select evidence | Claim | Primary evidence | Corroboration | |---|---|---| | Repo behavior | file plus test or command | history when provenance matters | | API, flag or version | installed manifest plus official docs for that version | release notes | | Best approach | one or two concrete alternatives measured in this repo | recognized reference | | External fact | primary source | two independent sources | If an authoritative source is inaccessible, do not substitute training memory; leave the claim undetermined. ## 3. Seek contrary evidence first - Look for the failing case, opposite passage, version mismatch or boundary that refutes the claim. - For a proposed solution, identify what each alternative does better. - Record every finding as `path:line`, URL or literal command output. ## 4. Seek favorable evidence Only after the contrary pass, collect evidence that sustains the claim using the same source standard. ## 5. Issue the verdict ```text Claim: Verdict: CONFIRMED | REFUTED | PARTIAL | UNDETERMINED Confidence: NN% Against: For: Alternatives: