Deep Review Plan

Deep Plan Review Skill

Run a structured, multi-agent review of an implementation plan before execution.

Scope Guard

This skill reviews plan files only (.md files in .claude/plans/, docs/plans/, or a plan presented in the current conversation). It does NOT review:

• Source code (.ts, .tsx) — use /deep-review instead
• Config files, skills, or documentation

If invoked and there is no plan to review, ask the user which plan file to review and stop.

Steps

1. Identify the plan: Check for a plan file referenced in the current conversation. If none, run ls -t ~/.claude/plans/ | head -5 and ls -t docs/plans/ | head -5 to find recent plans. Ask the user to confirm which plan to review. Read the full plan file.

2. Read project rules: Read /Users/tj/TJDevelopment/fireplanner/CLAUDE.md to have the full project rules in context for compliance checking.

3. Launch 4 review agents in parallel (single message, 4 tool calls):

   Agent 1 — Code Architect (subagent_type: feature-dev:code-architect): Review the plan for architectural soundness:

   • Does the plan target the correct files and modules?
   • Are the proposed abstractions appropriate (not over-engineered)?
   • Does the file organization follow CLAUDE.md rules (pure functions in lib/, hooks only if they call React hooks, data in lib/data/)?
   • Does the plan reuse existing shared helpers or duplicate logic?
   • Are there hidden dependencies between plan steps that aren't acknowledged?
   • Does it respect the two computation contexts (nominal vs real dollar basis)?

   Agent 2 — Feasibility Reviewer (subagent_type: feature-dev:code-reviewer): Review the plan for correctness and feasibility:

   • Do the referenced files, functions, and APIs actually exist in the codebase?
   • Are the proposed function signatures compatible with existing callers?
   • Does the plan violate any CLAUDE.md "Do Not" rules?
   • Are there any dollar basis mixing risks in the proposed calculations?
   • Is the scope realistic (does the plan underestimate complexity)?
   • Are edge cases acknowledged (empty states, zero values, missing data)?

   Agent 3 — Codex MCP (use mcp__codex-cli__codex tool): Send the plan content and ask codex to review for issues. Prompt: "Review this implementation plan for a Singapore FIRE retirement planner. Check for: logical flaws in the approach, missing steps, incorrect assumptions about existing code, scope creep, and any steps that contradict each other. The plan: [plan content]"

   Agent 4 — Gemini (use mcp__gemini-cli__ask-gemini tool, model: gemini-3-flash-preview): Send the plan content to Gemini for an independent review. Prompt: "Review this implementation plan for a TypeScript/React retirement planner app. Look for: logical flaws, missing steps, contradictions between steps, unrealistic scope estimates, and assumptions that might not hold. Also check if any step could introduce bugs or regressions in adjacent features. The plan: [plan content]"

4. Consolidate findings: Collect all 4 agent results, deduplicate, and present as a severity-ranked list:

   • BLOCKER: Plan violates CLAUDE.md rules, targets nonexistent APIs, or has logical flaws
   • WARNING: Missing edge cases, hidden dependencies, scope underestimated
   • SUGGESTION: Alternative approaches, simplification opportunities

5. Present revised plan if blockers were found. Otherwise confirm the plan is ready for execution and ask the user how to proceed.

Source: RemarkRemedy/fireplanner — distributed by TomeVault.