Learn · Prompt engineering & structured output · CCAR-F task 4.6

Independent and multi-pass review

A model checking its own work tends to agree with itself. Use a fresh reviewer, per-file passes plus an integration pass, and narrow checklists.

Key ideas

• Self-review in the same conversation shares the author's assumptions and tends to confirm them.
• An independent instance that sees only the artifact and the criteria catches more than "double-check your work" or extra thinking.
• For large PRs: per-file passes for focus, then an integration pass for cross-file issues.
• Specialised passes (security, performance, tests) with narrow checklists reduce false positives.
• Voting (best-of-N majority) for costly classifications; evaluator-optimizer loops need a stop condition.
• Give each finding its own confidence and route low-confidence ones to humans.

Worked example: Reviewing a 14-file PR

Run one reviewer subagent per file with the team's checklist, collect their findings, then run one integration pass that receives the diff summary and the per-file findings and looks only for cross-file problems such as a changed function signature with stale call sites.

Exam traps

• Asking the same conversation to "review your own code carefully".
• Reviewing a very large PR in a single pass and expecting thorough coverage.
• Giving the reviewer the author's reasoning, which anchors it on the same conclusions.
• An evaluator-optimizer loop with no stop condition.

Official docs

• Building effective agents
• Subagents in Claude Code

Practise CCAR-F questions · Next lesson: Context is a budget, not a bucket