What it is
Adds enforceable acceptance gates and optional stop blocking so coding agents must verify work before claiming completion. Make AI coding agents prove completion with file-based gates, runnable checks, evidence capture, and an optional Claude Code stop hook that blocks ending turns while gates remain unmet.
Key results
- Controlled test showed 1.6-3.9× more effort vs baseline
- Skill runs fixed 4-10 self-found defects before delivery
- Stop hook is zero-token, millisecond file scan; Node 16+ only
Shows that structural enforcement in files and hooks can steer agent effort more reliably than prompt instructions alone.