Spec Driven Agent Testing

Vibe coding is fast until the agent tells you everything passes and it does not. Coding agents quietly weaken tests to go green, fill gaps with "reasonable" defaults, read the exit code of the wrong command, and report success on work nobody checked. Spec-driven tools such as Spec Kit, Kiro and OpenSpec help you write the spec. This skill makes sure the code actually meets it. It is the quality layer that sits on top of any spec workflow, or works on its own. What your agent learns - Brief before building: correct behaviour, non-negotiable rules, acceptance checks and a do-not-build list, all written before any code - Make rules structural: put them in types and constructors so the wrong thing cannot be built, instead of hoping review catches it - Refuse, don't guess: stop and name the gap instead of inventing a default - Prove every gate: mutation-test each check so you know it can fail, read real exit codes, never trust a PASS that has never failed - Never weaken a test to go green - Review what comes back: re-verify claims and check the arithmetic in the agent's own summary Every rule comes with the real production defect it caught, eight in total. For developers, data scientists and researchers who let Claude Code or Codex write real code and need it to survive production, code review or peer review. One SKILL.md, any language, no framework lock-in.