What problem does it solve? AI-generated answers about completed work often claim success without running anything, invent file paths and API names, quote numbers without denominators, and bury blockers under filler. Readers act on these answers and lose time or ship defects. This Skill forces every claim in a report, review, or finding to be evidenced, quantified, or explicitly marked unverified. ## Core Features & Use Cases - Evidence ladder enforcement: Ranks every claim from bare assertion (rung 0) to before-and-after command output (rung 5), requiring unverified claims to name their rung instead of rounding up. - 50-pattern detection catalog: Detects unearned completions, fabricated APIs and terminal output, findings without file:line locations, percentages without denominators, buried blockers, and effort narration, each ranked into three severity tiers. - Claim-to-evidence lookup: Maps recurring QA claims (fixed, tests pass, coverage is N%, flake is fixed, no regression, ready to release) to the exact commands and outputs required to support them. - Use Case: Before sending a bug-fix report, run the audit checklist against the draft to catch a "tests pass" claim that only ran a partial suite, then rewrite it with the actual runner summary line and the skipped tests named. ## Quick Start Audit the answer you are about to send about the fix you just made, flag every claim that was not actually run, and rewrite it with real command output and file:line references.
