how-to-code-review

name	how-to-code-review
description	When the user asks for a code review, PR review, or critical look at recent changes, consult this skill to decide whether to delegate to the code-review subagent and how to craft the prompt.
user-invocable	false

Handling code review requests

Delegate to the code-review subagent (Agent tool, subagent_type: code-review) instead of reading diffs and writing the review inline. Review runs in a forked Opus context — diff reading, file expansion, and reasoning stay out of the main thread; only the report comes back.

When to delegate vs. handle inline

Delegate when:

• The user asks for a code review, PR review, or "look critically" at recent changes.
• The user asks "what's wrong with X" or "find issues in Y" about a defined scope.
• A natural review point has been reached (just-finished feature, pre-commit check).

Handle inline (do not delegate) when:

• The user wants a quick sanity check on a single small snippet visible in conversation. Delegating costs more than reading three lines.
• The user is asking how to fix a specific known issue, not asking for issue discovery.
• The user wants a discussion or back-and-forth about a design choice — that is conversation, not review.

Crafting the delegation prompt

The subagent's default scope is the working tree against the merge base with main. Override only when the user names something else.

Examples:

• User: "review my changes" → Delegate with empty prompt. Default scope handles it.
• User: "code review the auth refactor" → "Review the auth refactor in the current working tree."
• User: "review PR #482" → "Review PR #482."
• User: "look at src/api/handlers.go" → "Review src/api/handlers.go."
• User: "review the last 3 commits" → "Review the commit range HEAD~3..HEAD."
• User: "security-focused review" → "Review the current working tree with emphasis on security (auth, input validation, secret handling, injection)."

Anti-patterns

• Reading the diff yourself and writing the review in the main session. The subagent owns this; doing it inline burns the main context on diff content.
• Asking the subagent for full code blocks or diffs in its report. It returns findings keyed to file:line — the code is in the working tree.
• Delegating mid-discussion before the user has actually asked for review. "Should I refactor X?" is not a review request.