Code Review Skill | Agent Skills

Code Review

Review changes from a starting commit hash to HEAD, analyzing each commit individually like a human code reviewer would on a merge request.

Usage

Provide a commit hash to start the review from. The review will analyze all commits from that hash to HEAD.

Review Principles

Follow Google's Code Review Standards:

Design: Well-designed and appropriate for the system?
Functionality: Behaves as intended? Good for users?
Complexity: Can others understand and use this easily?
Tests: Correct and well-designed automated tests?
Naming: Clear variable, function, class names?
Comments: Clear, useful, explain why not what?
Style: Follows style guidelines?

Philosophy: Code should improve overall code health, even if not perfect. Balance forward progress with quality.

Important Guidelines

Big picture first: Understand the commit's purpose before examining specific code
Context matters: A line change only makes sense when you understand why it changed
Cross-file awareness: Changes in one file often relate to changes in others
Look for what's missing: Sometimes the issue is what WASN'T changed
Be specific: Reference exact line numbers and file paths
Be actionable: Each issue should have clear description and optional suggestion
No false positives: Only flag genuine issues matching the rules

Workflow

Step 1: Prepare Review Data

Run the CLI to collect commit metadata:

python3 ./scripts/cli.py prepare "<provided-commit-hash>"

Output is in TOON format (Token-Oriented Object Notation) for efficient LLM processing:

commits[2]{hash,author,date,subject,body,filesChanged}:
  abc1234...,Alice,2025-01-15,Add new feature,null,3
  def5678...,Bob,2025-01-14,Fix bug in parser,null,1
commitList[2]: abc1234...,def5678...
branch: feature-branch
commitRange: abc123^..HEAD
totalCommits: 2

Parse the commits array from the CLI output and display a table:

| # | Commit | Message | Files | | --- | ------- | ----------------- | ----- | | 1 | abc1234 | Add new feature | 3 | | 2 | def5678 | Fix bug in parser | 1 |

Use AskUserQuestion to ask which commits to review:

Option 1: "Review all" - review all commits in the table
Option 2: Free input for specific indices (e.g., "1,3,5" or "2-4")

Step 2: Load Review Rules

Read all rules from references/:

Step 3: Review Each Commit

For each commit in commitList:

Get the diff using python3 ./scripts/cli.py get-diff <commit-hash>
Apply the loaded rules to the changes
Record issues found

Issue Structure:

issues[2]{file,startLine,endLine,category,icon,severity,description,suggestion}:
  path/to/file.tsx,42,45,AI Slop,🧹,medium,Brief actionable description,Remove unnecessary comment
  src/utils/helper.ts,18,22,Component Reuse,♻️,low,Existing utility available,Use shared/utils

Severity Guidelines:

high: Bugs, security issues, broken functionality
medium: Design problems, complexity issues, missing tests
low: Style issues, suggestions, minor improvements

Step 4: Aggregate and Deduplicate

After reviewing all commits:

Deduplicate issues (same file + line range + category)
Calculate totals by category and severity

Step 5: Generate Summary

Read assets/summary-template.md and fill in the values:

| Variable | Description | | ------------------- | ------------------------------------------------- | | {commitRange} | Git commit range (e.g., abc123^..HEAD) | | {commits} | Total number of commits reviewed | | {files} | Total number of files changed | | {issues} | Total number of issues found | | {humanReviewTime} | Estimated human review time (2 min per 100 lines) | | {dedupLine} | Deduplication info or empty string | | {topIssues} | List of top issue categories with counts | | {recommendations} | List of key actionable recommendations |

Output the formatted summary.

Agent Skills: Code Review

Install this agent skill to your local

Skill Files