samchang72/dual-code-review

Dual Code Review Protocol

This skill guides you through performing a "Dual Code Review" where you act as the orchestrator to combine your own analysis with an independent review from the claude CLI.

Workflow

1. Preparation

1. Identify the scope of the review (e.g., git diff --staged, specific files, or a PR).
2. Constraint: Ensure the content to be sent to Claude CLI does not exceed context limits (approx. 200k tokens, but prefer smaller chunks for CLI speed).

2. Dual Analysis Phase

You must perform these two checks in parallel (or sequentially, but independent of each other's initial bias).

A. Antigravity Analysis (Yourself)

• Analyzes the code for logic errors, security vulnerabilities, performance issues, and coding standards.
• Note down your findings before reading Claude's full output if possible, to maintain independence.

B. Claude CLI Analysis

Run the following command to get Claude's perspective. Note: Adjust valid --staged or specific file paths as needed.

# Template Command
(
  echo "You are a Senior Code Reviewer."
  echo "Please review the following code changes."
  echo "Focus on: Bugs, Security, Performance, and Readability."
  echo "Return your findings in a structured JSON format with 'severity' (High/Medium/Low), 'file', 'line', and 'message'."
  echo "--- CODE START ---"
  git diff --staged # OR cat file.ts
  echo "--- CODE END ---"
) | claude --print

3. Consensus & Synthesis Phase (The "Mutual Agreement")

You act as the Judge. You must evaluate Claude's findings against your own.

Evaluation Logic:

1. Common Findings: If both you and Claude find the same issue → Confirmed High Confidence.
2. Claude Only:
   • Verify the finding. Is it a hallucination? Is it valid?
   • If valid → Accept & Include (Acknowledge Claude found it).
   • If invalid → Discard (Optionally note in reasoning why).
3. Antigravity Only:
   • Double check your own finding.
   • If valid → Include.

4. Final Report Generation

Generate a Review Report Artifact (e.g., review_report.md or directly in chat if short).

Report Structure:

1. Summary: High-level status (Pass/Request Changes).
2. Consensus Highlights: Critical issues agreed upon by both agents.
3. Detailed Findings:
   • Grouped by File or Category.
   • Tag findings: [Double Verified], [Claude CLI Detected], [Antigravity Detected].
4. Actionable Advice: Steps to fix.

Triggering

Use this skill when the user asks for "Dual Review", "Double Check", "Mutual Review", or "Review Process".