version: "1.0.0" name: review description: Review completed work against success criteria and flag issues user-invocable: true disable-model-invocation: false argument-hint: [what to review - file, feature, or recent work]

Review

Evaluate completed work against its success criteria. Find problems before they compound.

Input

What to review: $ARGUMENTS

If no arguments, review whatever was most recently implemented in the conversation.

Process

Step 1: Pre-check

• If /verify hasn't been run on the target, suggest running it first — /verify catches completeness issues, /review catches quality issues. Different concerns.
• Read relevant COMP files (CLAUDE.md for conventions, MEMORY.md for gotchas)
• Identify the success criteria (from /plan-task output, PLAN.md, or conversation)

Step 2: Launch Review

Read ~/.claude/skills/review/review-dimensions.md for what to check based on the type of work (code, data, writing).

Launch the review using a fresh-context subagent (use the code-reviewer agent persona for code). Apply the strengthened review discipline (per CLAUDE.md "Fresh eyes for review"):

1. Spec-blind: Do NOT include the original spec, plan, or coder context in the reviewer's prompt. Pass the diff (or the artifact) and have the reviewer reason backward. The reviewer's lack of spec-knowledge is a feature — it surfaces unstated assumptions the spec hid.
2. Multiple uncorrelated passes: Launch the reviewer 2-3 times in parallel with the same prompt. Union their findings. If the same model produces overlapping findings, that's signal; if they find different things, the union is the value.
3. Treat-as-adversary: Frame the reviewer's findings as questions about the work, not defects to auto-fix. Pass each finding through "is this actually wrong, or is it the reviewer pattern-matching without context?" Keep the credible ones; reject the rest with reasoning.

Step 3: Present Findings

Read ~/.claude/skills/review/output-format.md for how to organize findings.

Step 4: Escalation Check

• For significant reviews, consider suggesting /dialectic-review for adversarial depth
• If the review surfaces architectural concerns, suggest /dialectic-review --premortem before proceeding with fixes

Rules

Read ~/.claude/skills/review/rules.md for review rules and gotchas.