What problem does it solve? AI agents produce code faster than humans can review it, and prose rules (CLAUDE.md, comments, docs) do not stop recurring failures like phantom wiring, invented API signatures, and silently-green tasks. This Skill turns quality into a mechanical validation loop instead of relying on someone remembering to review. ## Core Features & Use Cases - Non-eludible validation loop: three layers (pre-commit, blocking CI, pre-push) with rules that make gates actually able to fail, plus named escape hatches. - Ratchet baselines for dirty codebases: turn on a new check over existing violations by freezing known debt and failing only if it grows, indexed by stable identity rather than line numbers. - Self-verifying documentation: <!-- verify --> blocks let docs assert numbers (LOC, file counts, regex matches) that scripts/doc_verify.py recalculates and fails the build on drift. - Agent-specific failure modes: checklists and guards for invented API signatures, unverified subagent self-reports, phantom wiring, and silent-green periodic tasks. - Use Case: You are bootstrapping a project that Claude Code or similar agents will mostly write. Apply the adoption order: destructive-action guards first, then a minimal validation loop, measured lint debt, one project-specific check with a ratchet baseline, and verify blocks on the fastest-aging numbers. ## Quick Start Ask the agent to set up the agent-ready validation harness for this repository, starting with the destructive-action guards and a pre-commit plus CI loop that can genuinely fail.
