critic-overkilllisted

# ClaudeCritic Overkill Three expert-critic sub-agents. One score per round. Only convergent fixes get implemented. Loops until 99/100 or graceful plateau. This skill captures a pattern that emerged from real long-running polish sessions: dispatching multiple critics with distinct expertise produces more honest critique than a single reviewer (which tends to self-agree), and only implementing issues flagged by ≥2/3 critics filters out solo-taste noise. ## When to use - User asks to **refine**, **polish**, **perfect**, **harden** something - User sets a `/goal` containing phrases like "until critics give 99/100", "production-grade", "make it perfect" - User says "score-loop on Z", "iterate until done", "loop until you can't improve it" - User wants component-by-component or endpoint-by-endpoint refinement ## When NOT to use - User wants to **build something new** — use a builder skill, or just regular Claude Code - User wants to **find bugs / red-team** — use `adversarial-break` or similar - User wants to **audit current state without changing it** — use a code-review skill ## First-run initialization (mandatory) Before running any loop, check for `.critic-overkill/config.md` in the project root. **If missing**: run `INIT.md` — a short auto-detection pass that reads `CLAUDE.md`, `package.json` / `pyproject.toml` / `go.mod` / `Cargo.toml`, the README, and a sample of source files. Then write `.critic-overkill/config.md` containing: - `language`: detected commit-messag