Staff Review

Skills

The /staff-* skills that drive the review, and how to make them your own.

staff install writes nine skills to .agents/skills/ (the six below plus the three building blocks), symlinked into .claude/skills/ so Claude Code exposes them as slash commands. They’re just Markdown — edit them to encode your team’s standards.

SkillWhat it does
/staff-reviewReviews the active diff with a fan-out of sub-agents, verifies their findings, and leaves inline comments.
/staff-resolveWorks each open thread: fixes the code, documents it, or skips with a justification.
/staff-commentThe thin CLI wrapper the others use to post/edit/resolve comments.
/staff-documentImports a GitHub PR review comment (by URL) as a .staffreview/docs/ example.
/staff-docsMines past Staff Review diffs and GitHub PRs for recurring lessons and writes the best ones into .staffreview/docs/.
/staff-loopReviews (find → verify → post) and then resolves the diff in sub-agents, round after round, until it converges.

Three more are building blocks, not run directly: /staff-review-find (one find agent — reviews an assigned slice of the areas/docs and returns findings), /staff-review-verify (one verify agent — re-checks a batch of findings), and /staff-docs-scout (one scout agent — reads an assigned slice of history and returns candidate lessons). /staff-review and /staff-loop fan out to the first two, and /staff-docs fans out to the scouts, so the orchestrators never have to spawn a top-level skill as a sub-agent.

/staff-review

Conducts a staff-engineer-level review of the active diff as a multi-agent orchestration. It fans the work out across parallel sub-agents — each owning a slice of the ten review areas (correctness, edge cases, concurrency, security, data, interfaces, tests, consistency, readability, performance) and a slice of the review docs — so a large diff and a growing docs both stay fast. A second wave of agents then verifies every finding against the code to drop false positives before anything is posted. Survivors land as inline comments, each with a priority (P1 must-fix, P2 should-fix, P3 minor) so you can triage.

How many agents it fans out per phase is the reviewAgents setting (default 2, range 1–20, adjustable in the gear menu). You can also override it inline for a one-off — pass a bare number after the slug to tailor it to the change's size:

/staff-review main..WT        # uses the reviewAgents setting (default 2)
/staff-review main..WT 6      # fan out wider for a big diff

The review runs ~2× that many agents in total (a find wave, then a verify wave), which is why the cap is conservative.

/staff-resolve

Works through every open thread. For each one it picks exactly one action — fix the code, document it as a docs entry, or skip it with a justification — replies in-thread with what it did, then records the resolution. It does not commit; you review and commit the changes.

/staff-resolve

/staff-comment

The thin CLI wrapper the other skills use to add, edit, delete, list, and resolve comments via the staff CLI. You rarely invoke it directly.

/staff-document

Imports a GitHub PR review comment by URL and turns it into a structured .staffreview/docs/ example — a way to seed the docs from review wisdom that already lives on GitHub.

/staff-docs

Mines your review history — past Staff Review diffs and GitHub PRs (via gh) — for recurring, generalizable lessons worth keeping, then writes the best ones into .staffreview/docs/. It fans the reading out across parallel /staff-docs-scout sub-agents, ranks the survivors impact-first, surfaces the top candidates for you to flag, and documents the ones you keep. The scout fan-out width is the docsAgents setting (default 5, range 1–20, adjustable in the gear menu).

/staff-docs

/staff-loop

Each round runs a full multi-agent review (fanning out the same /staff-review-find and /staff-review-verify sub-agents that /staff-review uses) and then a /staff-resolve sub-agent, round after round, until a fresh review finds nothing new or it hits the loopMaxRounds cap (default 5). It runs the review itself rather than calling /staff-review, so it never nests orchestrators. Point it at a working-tree diff so each round’s fixes feed the next review.

/staff-loop main..WT

Make the skills yours

They’re Markdown in .agents/skills/. Add your conventions, your “don’t do X here,” your preferred test framework. A review is only as good as the standard you hand it.

On this page