build-guardrail

What problem does it solve? Written coding rules decay because nothing enforces them mechanically. This Skill turns a rule into a test that turns red on violations, with a baseline and a ratchet that only allows the violation count to decrease, plus a self-falsification statement so the guardrail can say when it itself is broken. ## Core Features & Use Cases - Rule-to-check conversion: Picks one countable quantity per rule, measures from actually loaded runtime state rather than static config, and classifies conservatively (undecidable items never count as safe). - Anchor and injection discipline: Requires self-falsification statements written before measurement, synthetic-input anchors instead of real-world defect counts, and two-direction injection tests proving the check both reports bad input and stays quiet on good input. - Baseline and ratchet workflow: Generates baselines via vitest, enforces named-item reporting on regressions, distinguishes hard-zero rules from ratchet rules, and records reviewed decisions in a decisions JSON with collision-resistant keys. - Use Case: A team wants to enforce "no cross-layer direct calls" in a TypeScript codebase. The Skill guides writing an audit test that scans loaded modules, fails red on first run, names each violation, then locks a baseline so the count can only go down. ## Quick Start Ask the AI to build a guardrail for the rule you want enforced, naming the quantity to measure and the test file where the check should live.