/critic — Plan Review Wrapper

Adversarial second-opinion pass on a plan or proposal before execution begins.

When to invoke

• Plan file written, before ExitPlanMode approval
• Long autopilot/ralph prompt drafted, before launch
• User asks “is this plan any good”
• After a deep-interview produced a spec, before implementation

When NOT to invoke

• Trivial single-file changes
• Bug fixes where the fix is mechanical
• During implementation (use code-reviewer instead)

Workflow

1. Identify target:
   • If active plan file exists at ~/.claude/plans/*.md (latest mtime) → use it
   • Else use the last assistant message in conversation
   • User can pass explicit path: /critic <path>

2. Spawn oh-my-claudecode:critic agent (model=opus) with prompt:

   Review the attached plan. Find 3-10 specific issues across these axes:
   - Scope creep (does it sneak in unrelated changes?)
   - Symptom vs root cause (is it patching a symptom?)
   - Missing tests / verification gaps
   - Fabricated facts (file paths, function names, APIs that may not exist)
   - Unverified assumptions about codebase state, user intent, library behavior
   - Sequencing risks (steps that depend on prior unproven steps)
   
   Format each finding:
   - [SEVERITY: HIGH|MED|LOW] <one-line summary>
     - Location: <plan section or line>
     - Why it matters: <impact if not fixed>
     - Recommended fix: <concrete change>
   
   Max 10 findings. Return only the findings. No preamble, no summary.
   

3. Surface findings verbatim to user. Do NOT auto-edit the plan.

Hard Rules

• Critic agent is READ-ONLY. No file edits.
• Findings include file:line where applicable.
• Severity is honest, no padding to hit a count.
• If plan is genuinely solid: return 0–2 findings + explicit “no major concerns” line. Don’t manufacture issues.

Triggers

/critic, “review my plan”, “plan critique”, “adversarial review”, “second opinion on plan”