Code Review Guidelines

Purpose

Code review catches bugs, ensures maintainability, and builds team knowledge. This skill provides a structured, objective, empathetic framework for reviewing code without causing friction. The goal is to improve the codebase while growing the team.

When to use

• Evaluating a Pull Request or Merge Request
• Providing constructive feedback to peers
• Ensuring code quality and architectural standards are met
• Building new team members' skills through feedback

When NOT to use

• Nitpicking personal coding style (use linters/formatters)
• Blocking on subjective preferences
• Reviewing without understanding context

Inputs required

• Pull request with code changes
• PR description explaining the why
• Linked ticket with context
• Test coverage (unit/integration tests)

Workflow

1. Understand the Goal: Read the PR description and linked ticket FIRST. Understand WHY the change exists before reading code.
2. Review Architecture: Check structural integrity and design. Does it fit the codebase? Are new dependencies justified? Is it the right abstraction level?
3. Review Logic: Check for correct error handling, edge cases, off-by-one errors, and security implications.
4. Check Tests: Verify new logic is covered by unit/integration tests that test actual behavior (not just code coverage).
5. Consider Performance: Check for obvious performance issues (N+1 queries, memory leaks, synchronous blocking in async code).
6. Provide Feedback: Leave comments that explain the WHY. Differentiate between blocking requests and suggestions (use "Nit:" or "Note:").
7. Be Constructive: Phrase feedback as questions when possible ("Have you considered...?"). Assume positive intent.

Rules

• MUST assume positive intent; critique the code, not the author
• MUST automate style and formatting checks via Linters/Formatters (do NOT argue over spacing in PR)
• MUST approve immediately if code improves the codebase, even if not perfect
• MUST provide context and links when referring to standards or patterns
• MUST distinguish between blocking requests and optional suggestions
• MUST NOT block on personal coding style preferences
• MUST NOT require unrelated changes (keep scope focused)

Anti-patterns

• Gatekeeping: Blocking PRs over personal style preferences or "I would have done it differently"
• Rubber Stamping: Approving large PRs without reading or thinking
• Scope Creep: Asking the author to fix unrelated legacy code nearby
• Vague Comments: "This is bad" or "Improve this" without explanation
• Late Requests: Requesting major architectural changes after code is already implemented
• Tone Issues: Comments that feel dismissive or condescending

Failure conditions

• PR blocked without clear explanation
• Security or correctness issue missed
• No tests for new logic
• Feedback based on personal preference, not code quality
• Author feels disrespected or attacked

Validation checklist

• [ ] PR description explains the why (not just what)
• [ ] Tests cover new logic
• [ ] No obvious performance issues
• [ ] Error handling is correct
• [ ] Edge cases are considered
• [ ] Security implications reviewed (if applicable)
• [ ] Code follows established patterns in the codebase
• [ ] Dependencies are justified
• [ ] Comments explain complex logic
• [ ] No scope creep (focused on this change)

Output format

• Feedback structure: Problem + context + suggestion
• Tone: Constructive, curious, respectful
• Clarity: Blocking requests vs. optional suggestions clearly marked
• Actionability: Comments are specific and actionable
• Decision: Clear approval, changes requested, or comment

Security considerations

• Security issues MUST be flagged as blocking
• SQL injection, XSS, authentication bypasses MUST be caught
• Secrets/credentials MUST NEVER be committed (catch before merge)
• Permission checks MUST be enforced
• Data exposure MUST be reviewed

Agent execution notes

• Agent MAY: Leave constructive feedback, ask clarifying questions, suggest improvements
• Agent MUST NEVER: Block on style/preference, be dismissive, miss security issues, approve without reading
• Agent MUST ASK: When context is missing, when architectural impact is unclear
• Agent MUST VALIDATE: Tests are present, no security holes, code improves codebase

Example

❌ Anti-pattern (Vague, dismissive, scope creep):

"This code is bad"
"Why are you doing it this way?"
"Also, can you fix the bug in the file next to this?"
"Blocked - needs improvement"
"LGTM" (without reading)

✅ Correct pattern (Constructive, specific, respectful):

✅ Good catch on the edge case - this prevents the race condition we had in production.

💭 Question: In the validateEmail function, what happens if the email service is down?
   Have you considered adding a timeout or fallback behavior?

🔒 Security: Double-check that user input is sanitized before the database query.

📝 Nit: This comment could be clearer - what does "process the thing" mean?

✅ The test coverage looks good. I especially like the edge case tests.

🚀 Approved! This is a clear improvement to the codebase. Nice work.