← ClaudeAtlas

pr-reviewlisted

Deep review of someone else's pull request in an isolated worktree — hunts the failures that automated review misses (broken error paths, infinite loops, races, blast radius, claim-vs-code drift), deduplicates against the bot's comments, and gives a short verdict in chat. Read-only; never edits, pushes or posts.
grinchenkoedu/claude-skills · ★ 0 · Code & Development · score 75
Install: claude install-skill grinchenkoedu/claude-skills
# /pr-review — review someone else's pull request The last check before a change lands. By the time this runs, an automated reviewer has usually already commented and CI has usually already passed — so this skill deliberately does **not** repeat that work. It looks for what those layers structurally cannot see. Read-only: it never edits, commits, pushes, or posts a comment. You paste what you want to say yourself. For your *own* changes before pushing, use `/review` instead. ## What to skip, and what to hunt **Skip** — formatting, import order, naming preferences, docblock style, and anything the automated reviewer already commented on. Mention a basic only when it is the visible tip of a real defect: report the wrong value it produces, not the style rule it breaks. **Hunt** — the things a diff-shaped reviewer misses: - **Failure paths** — what happens when the happy path breaks halfway. An exception between two writes, a transaction left open, a half-updated record, a retry that duplicates work. - **Loops that may not end** — walk every new loop, search or cursor against hostile data: gaps in ids, empty sets, duplicates, nulls. Mocked tests prove nothing here; check the algorithm itself. - **Races and double-submits** — check-then-act, a form submitted twice, a scheduled task overlapping its previous run, a unique constraint that deduplicates the *row* but not the *action*. - **Data safety** — see the data-safety rules in `reference/repo-profile.md`. Unbound