judgelisted

# /judge — LLM-as-Judge Artifact Evaluation Evaluates key artifacts against criteria using an isolated evaluator context. The evaluator has no access to the generator's reasoning — only the artifact and the rubric. ## Usage ``` /judge <artifact-type> [artifact-path] ``` Artifact types: - `product-brief` — evaluates `product_brief.md` - `contract` — evaluates `contract.json` completeness and quality - `feature` — evaluates completed feature against `contract.json` criteria - `design` — evaluates wireframe against `design-contract.json` `plan` is intentionally not a generic judge type. `/pm-review` is the sole Phase 2 plan gate and writes `pm-review.json`; do not create a second verdict for the same transition. ## Isolation principle The judge **must not** be the same agent that created the artifact. Use a separate context: ```bash # Claude Code: fork a new agent for evaluation Task({ subagent_type: "general-purpose", prompt: "You are an evaluator agent. Your only job is to score artifacts against rubrics..." }) ``` ## Instructions (evaluator agent) ### System prompt for evaluator ```xml <role> You are a strict evaluator with 10+ years of product and engineering review experience. You were NOT involved in creating the artifact. You have no bias toward approving it. Your job: find gaps, not validate effort. </role> <constraints> <required> - Score each criterion 0.0–1.0 with specific evidence - List all issues, even minor ones - Verdict: PASS (all mu