What problem does it solve? Manually verifying that a Claude plugin injects the right skills into realistic projects is slow and error-prone. This Skill automates the full loop: creating test projects, exercising skill injection via claude --print, launching dev servers, verifying they respond, analyzing conversation logs, and producing actionable improvement reports. ## Core Features & Use Cases - Four-Stage Pipeline: Chains runner, verify, analyze, and report stages sequentially, aborting on failure and logging NDJSON events to events.jsonl. - Structured Contracts: Uses run-manifest.json to correlate JSONL sessions with projects precisely, and emits machine-readable report.json with verdicts, gaps, and suggested frontmatter patterns. - Self-Improvement Loop: Supports overnight automation with cooldown loops so report.json verdicts can be compared across runs, trending from fail to pass. - Use Case: Run the quick mode (--quick, 3 projects) before committing a change to the vercel-plugin to confirm expected skills like auth, payments, or cron-jobs are still injected correctly. ## Quick Start Ask the AI to run the benchmark-e2e pipeline in quick mode to verify skill injection across the first three test projects.
