| name | self-review |
| description | Spin up a dedicated review sub-agent to critique Claude's own pending changes or recent commits, following the adjustable methodology in this skill. Use when the user asks to review your changes, review the diff, review the last commit(s), do a self-review, or types /self-review. Optional argument = the scope to review (e.g. a commit ref, a range, "last 2 commits", "staged"). |
Self-review
Delegate a critical review of the current changes to a fresh sub-agent, so the
review isn't biased by the context that produced the code, then report its findings
back to the user.
1. Determine the review scope
Use the argument passed to the skill if there is one. Otherwise auto-detect, in this order:
- Uncommitted work — if
git status --porcelain is non-empty, review the working
tree: staged + unstaged + untracked. Reproduce it with git diff HEAD and list
untracked files via git status --porcelain.
- Branch commits — otherwise review commits on this branch that aren't on
main:
git log --oneline main..HEAD and git diff main...HEAD.
2. Spawn the review sub-agent
Do not review the code yourself in the main thread — the whole point is a fresh pair of eyes. Launch one sub-agent with the Agent tool (subagent_type: general-purpose,
description like "Review current diff"). Its prompt must contain:
- The exact scope you resolved (the git command(s) that reproduce the diff, or the branch/range).
- This instruction, verbatim:
Read .claude/skills/self-review/methodology.md and follow it exactly. Review only
the changes in the scope below. Do not modify any files.
- The scope details.
3. Relay the findings
Summarize the sub-agent's report for the user: lead with the verdict and any blocking
issues, then the rest grouped by severity, keeping file:line references clickable.
Don't silently drop findings. Offer to fix them, but don't start fixing unless asked.