What problem does it solve? AI assistants routinely claim a change is done, fixed, or working based on their own assumptions, a green build, or a passing lint run — and defects ship. This Skill forces an independent verification step before any such claim is made, so unverified work never gets reported as finished. ## Core Features & Use Cases - Fresh adversarial verifier subagent: Spawns a verifier with zero knowledge of what was changed or why, receiving only the verbatim requirement and the changed-file list from git status / git diff. - Live-evidence bar: Requires real HTTP calls, real browser runs via Playwright tools, real data queries, and host commands — mocked unit tests and passing builds are automatic FAILs. - Four named failure-mode sweep: Explicitly checks partial coverage, over-application, deployment-present-vs-behaviour-verified, and happy-path-only testing, each reported as PASS, FAIL, or UNVERIFIED. - Use Case: After fixing a login bug, instead of telling the developer "fixed", the assistant hands the original requirement and changed files to a fresh verifier that starts the app, drives the real login flow in a browser, probes edge cases, and returns a numbered evidence log. ## Quick Start Before claiming a change is done, ask the assistant to verify this change with a fresh adversarial verifier using only the requirement and the changed-file list.
