Claude Code Review: Self-Review vs an Independent Second Pass
Claude can review a local diff. The design question is whether the author should also own the final claim that the task is complete.
Use Claude Code’s local review for fast feedback inside the author session. Use managed PR review when the team wants GitHub-centered coverage. Add an independent finish pass when the author’s task interpretation, test selection, and completion claim should be challenged by a separate role before the work moves on.
Claude Code now has more than one review path
The phrase “Claude Code review” can refer to a local command that reviews a diff, a deeper cloud review, or managed review attached to GitHub pull requests. Anthropic’s current documentation separates those surfaces and their scope. Start by deciding whether you want iteration help, a team PR workflow, or a completion decision.
- 1Is the code still changing?Use Claude’s local review for fast feedback while the author is iterating.
- 2Does the team need shared PR visibility?Use the managed pull-request path when comments, permissions, and repository coverage are the main requirement.
- 3Is this a consequential completion claim?Add a separate finish pass with a fresh reconstruction, executable evidence, and explicit limitations.
| Path | Use it for | Boundary |
|---|---|---|
| Claude local review | Fast feedback in the current coding session | Same product and author context can influence the review. |
| Claude managed PR review | GitHub comments and organization-wide pull-request coverage | Requires a pushed pull request and the managed service. |
| Critique sidecar | An independent finish result for local working-tree changes | It adds another pass; it does not replace CI or human ownership. |
| CritiqueCode | A separate author agent with review enforced by its controller | Choose it as the author, not as a plugin layered on Claude Code. |
Keep Claude Code when it is already the right author
There is no reason to replace a productive author loop merely to obtain an independent review. Keep Claude Code, install the critique sidecar, and let the author call the finish pass once a meaningful unit of work is ready. The result comes back to the same conversation as structured evidence and limitations.
Add the sidecar without replacing the author
npm install --global @critiquedotsh/cli
critique login
critique integrate --agent allWhen to choose CritiqueCode instead
Choose CritiqueCode only when you want Critique to author the change. It runs an interactive coding session, then the controller owns the review and promotion boundary. It is @critiquedotsh/harness, binary critique-code; the sidecar is @critiquedotsh/cli, binary critique. Mixing those jobs makes adoption harder and the comparison less honest.
A practical setup is to keep the author’s normal Claude Code commands unchanged, call the sidecar only at a meaningful finish boundary, and save the result with the change record. That preserves a quick inner loop while making the final claim auditable by someone who did not write the patch.