Immutable. This exact content is served forever at /api/v1/blob/4dccbb11eca57dbe.
--- name: judge description: Use when the user wants an independent expert review of work done in this conversation before accepting or extending it. argument-hint: "[description of what to review, or omit for recent work]" --- subject = $ARGUMENTS No subject → review most recent substantive work in this conversation; identify from context, confirm with user before spawning. Subject references a path → read enough to know full scope before dispatching. **Why this exists:** the engineer who built something is the worst judge of it — they see what they intended, not what's there. Independent reviewers with clean context catch approach-level errors and missed alternatives that the original author rationalizes away. This skill produces that signal without the cost of a human code review. ## Role You coordinate; teammates judge. Their value is the clean context window — they didn't build this, so they aren't biased by it. Don't run the review yourself; dispatch + synthesize. ## Scope Not `/audit-necessity` (should it exist). Given we're building it, **did we build it the way an expert would?** Approach selection, architecture fitness, idiom correctness, trade-off awareness, missed alternatives, domain-standard solutions, proportionality. ## Teammates Spawn three clean-context reviewers in parallel, read-only and analysis-only. Domain, proportionality, and alternative-path coverage is the minimum panel. | Teammate | Agent file | Lens | |----------|-----------|------| | Domain Expert | `domain-expert.md` | Would a senior specialist in this exact domain do it this way? | | Pragmatist | `pragmatist.md` | Is this the most direct path to the goal? | | Alt-Path | `alt-path.md` | What fundamentally different approaches did we not consider? | Each gets: subject description, all relevant file paths / code / context, and the project's stack + conventions (detect from codebase). Each reports **everything its lens surfaces**, unfiltered — the credibility pass below is yours, and it only works on a complete list. A teammate that pre-filters hands you a short list you can't distinguish from a thorough one. ## Synthesis ### Credibility filter This is the filter stage; teammates do not run it. Drop any finding that fails one of: 1. **Substantiated** — cites specific code/decision/pattern, not "generally speaking". 2. **Actionable** — proposes a concrete alternative, not just criticism. 3. **Trade-off honest** — states the alternative's costs too. 4. **Calibrated** — separates "this is wrong" from "valid but here's another" from "fine, style preference". Overclaiming is this skill's #1 failure mode. Also drop: style preferences dressed as expertise, findings where the teammate misread the constraints. ### Convergence 2+ teammates independently flag the same concern → elevate it. Teammates contradict → present both with reasoning, don't pick a winner. ### Verdict scale | Verdict | Meaning | |---------|---------| | **EXPERT-GRADE** | A domain expert would recognize this as their own work. Style nits at most. | | **SOLID** | Sound approach. Real improvements found, no fundamental issues. | | **RETHINK** | Functional, but an expert would take a meaningfully different approach. | | **RED FLAG** | Fundamental approach issue. Specific alternative(s) strongly recommended. | ### Report 1. **Verdict** — one word + one-sentence justification 2. **What's strong** — what teammates validated (criticism-only reports read as dishonest and lose trust) 3. **Findings** — grouped by importance. Each: concern, evidence, proposed alternative, trade-off of alternative, source teammate(s) 4. **If we could start over** — single highest-leverage change, if any **Stop after the report. Do not implement changes unless asked.**