PR review assistants
Help on both sides of the code review
Now that AI helps developers push code faster, there is more to review than ever: code review is the new bottleneck.
These skills assist in both directions: when others leave feedback on your PRs, and when you review someone else's code. Unlike fully automated review bots, they work with you, locally: the agent suggests, you decide, and nothing reaches your teammates without passing through your hands.
fetch-pr-review
Give it a PR link and it downloads all the feedback your PR received into a self-contained PR-REVIEW.md. Works with GitHub, Azure DevOps or whatever platform your agent can access (private repos need the right access tokens or MCP server configured). Fetching is all it does: the thinking happens in a fresh session.
refine-pr-review
Goes through that file with you, comment by comment. For each one the agent reads the code the comment is about, then says what it would do: address it, address part of it, or push back, with its reasoning and a reply you can post. The debatable ones come one at a time, the obvious ones in a single batch, so you spend your time where it matters. You end up with two files: an ANSWERS.md with a reply per comment, to paste once your fixes are in, and a REQUIREMENTS.md with the changes you accepted, ready to apply right away when the fixes are small, or to hand to /create-implementation-plan when they need a plan. Comments that point at a habit, not a one off slip, end up as lessons at the bottom of ANSWERS.md, ready for /self-improve so your agent doesn't repeat them on the next task. Your PR is never touched.
review-code-assistant
A quick pass that prepares a review with you, locally in your terminal, and posts nothing. Point it at a PR link or a branch: it reads the same diff the platform shows, the PR intent and the project's convention docs, then suggests candidate comments in a human voice, ordered exactly like the diff so you can scroll the PR and the terminal side by side. A comment may only exist when it points to concrete evidence: an actual failure, a cited convention, a provable simplification. Zero comments is a valid outcome. You decide what to post.
maintainer-review
For when the merge is your call. Give it a link to someone else's PR and it reads everything first: the description, every linked issue and PR, and every comment, bot comments included. For each comment it checks two things: was it answered, and does it still apply to the latest code. A claim like "fixed in the latest push" is something to verify, never a fact, whoever wrote it.
Then it reviews the diff the same way /review-code-assistant does, and lets a sub-agent with fresh eyes double check its own findings before you see them. You get one verdict: does anything block the merge? Nothing gets posted, approved or pushed until you say so, and you see the wording first. If you ask, it can push a fix to the contributor's branch, without ever rewriting their history.