What problem does it solve? Editing a CLAUDE.md, skill, agent definition, or system prompt gives no proof the change actually works, since your current session already contains the change and any in-session check trivially passes. This Skill verifies prompt changes by spawning a fresh, clean-context Claude agent that re-reads the prompt from disk and observing whether the intended behavior emerges. ## Core Features & Use Cases - Clean-context evaluation: Runs claude -p sessions that load the real production prompt hierarchy, avoiding false results from in-session checks or under-applying subagents. - Sampled pass rates with A/B attribution: Runs 3-5 neutral trigger tasks, compares against a pre-change baseline, and reports a ✅ / ❌ / 🤔 verdict with a pass rate like 4/5. - Use Case: After adding a ✅/❌/🤔 status convention to your global CLAUDE.md and running home-manager switch, use this Skill to spawn five fresh Opus sessions with neutral tasks and confirm the convention is applied before calling the change done. ## Quick Start Verify my CLAUDE.md change works by running five fresh claude -p sessions with a neutral task and reporting the pass rate.
