review-pr
/review-pr — Deep GitHub PR Review
Reviews a remote GitHub PR with anti-slop filtering. Input: PR URL only.
Goal: produce an accurate, critical, actionable PR review that surfaces what a human reviewer should double-check — and filters out noise (style nitpicks, hallucinated references, duplicates, generic advice).
This skill assumes CodeRabbit is configured on the repo via .coderabbit.yaml (template at ~/.claude/skills/coderabbit-config/). CodeRabbit catches style + convention findings before this skill runs; /review-pr focuses on what only deep semantic + codebase-wide review can do.
Use AskUserQuestion for ALL user-facing decisions — stop-and-ask, cache replay, large-PR confirmation, self-review, post-review, post-failure, post-completion. Always cursor-selectable, never plain-text numbered lists.
Anti-patterns — NEVER do these:
- NEVER present choices as a numbered markdown list in terminal text. That short-circuits the tool call.
- NEVER end a response with
Would you like me to... ?followed by options in prose. - NEVER ask
What's next — do X or are we done?after a tool-call decision. - Self-test: if you catch yourself writing a sentence that asks the user to pick between 2+ labeled paths, STOP and use
AskUserQuestion.
Reference files (loaded on demand by subagent)
The Claude reviewer subagent loads these when relevant — keep them in mind so this file stays small:
More from bhagyamudgal/skills
fix-pr-review
Triage and fix CodeRabbit / review-pr findings on a GitHub PR, then reply + resolve conversations. Classifies each finding (FIX/DISMISS/DEFER/DISAGREE/NEEDS-INPUT), runs /done, posts specific replies. Use on CodeRabbit review URLs, PR URLs, or local review files. Pairs with /review-pr.
8harden-plan
Pre-code quality gate that runs /review-pr's anti-slop lens against a written plan BEFORE any code is written. Grounds the plan against the real codebase, runs 11 category checks (security, concurrency, round-trip, control-flow, error-handling, pattern-consistency, plus /review-pr's Q1-Q6), then grills the user one question at a time until the plan is hardened. Use when the user says "harden my plan", "check my plan", "grill my plan before I code", "lint this plan", or invokes `/harden-plan` explicitly. Also invoke proactively after `/superpowers:brainstorm` or `/grill-me` completes with a written plan and before any implementation begins. Do NOT invoke after coding has started — redirect to `/review-pr` / `/fix-pr-review` in that case.
8forge-plan
Use when starting any non-trivial feature or design task that needs the full idea-to-implementation pipeline. Also use when user says "forge", "forge plan", "full pipeline", or wants the complete AI development workflow before shipping.
6project-discovery
Deep project discovery and architecture planning. Use when starting a new project, when user says "new project", "let's build", or asks for help architecting/planning a non-trivial application.
6qa
Use when user says "qa", "test the flow", "QA this", "verify the UI", "test this page", "check if it works", "end-to-end check", or after completing code changes that affect UI and need manual verification replaced by automation.
5done
MANDATORY post-task verification. Run after EVERY task — no exceptions, no skipping, regardless of task size. Executes type-check, parallel code review, and code simplification in sequence.
5