build-looplisted

Autonomous generator-evaluator cycle that builds a feature against contract.json, with the evaluator launching the live app via Playwright MCP for grading. Generator and evaluator have separate context windows. Loops until pass, restart, or max iterations. Use after /contract is written, for new features (not bugfixes — those use /diagnose + /tdd). User triggers — /build-loop, "autonomous build", "let it cycle", "generator-evaluator", "GAN loop". Requires Playwright MCP installed and contract.json present.
createusernam/setup_project · ★ 0 · AI & Automation · score 68

Install: claude install-skill createusernam/setup_project

# /build-loop — Autonomous Generator-Evaluator Cycle ## Why this skill exists From the Anthropic Applied AI talk (May 2026): the load-bearing pattern for long-running coding agents is a **generator and evaluator in separate context windows**, grading against a **contract negotiated on disk before any code is written**. Three rules from the talk that this skill enforces: 1. **Evaluator never sees generator's reasoning or traces.** Sycophancy contagion otherwise. 2. **Evaluator launches the live app**, not reads diffs. Playwright/Chrome MCP clicks, types, screenshots, measures. 3. **Restart > patch on hard failures.** Use the contract's restart threshold instead of accumulating low-progress patches. ## Where this fits ``` /planning-with-files → /design-rubric (UI only) → /contract → /build-loop → /code-review-expert → ship ▲ you are here ``` For **bugfixes**: don't use this skill. Use `/diagnose` → `/tdd`. /build-loop is for greenfield/feature work with subjective quality dimensions (UI, design, originality). ## Prerequisites — HARD GATE `scripts/run.sh` enforces every item below before it creates mutable loop state. Static contract and attestation checks run in `validate-prerequisites.py`; the iteration contract has its dedicated validator; runtime reachability is represented only by trusted orchestrator environment handoffs; preflight and GRACE