Performs one in-depth multi-perspective code review of either a pull request or a commit range. Launches up to ELEVEN specialized parallel reviewer roles (AGENTS.md compliance, shallow bug scan, git history context, prior PR comments, in-file code comments, database / data-layer, OWASP Top 10 security, error handling, test coverage, headline-benefit / motivation delivery, and — unless `--skip-ticket` is passed — ticket intent compliance), then scores each finding 0–100 for confidence, filters anything below 70, and deduplicates. Returns the surviving findings. Never writes to GitHub. Used as one of two parallel review primitives (the other being `gh-style-review`) by `review-and-fix` (which spawns 3 of each per iteration and adds a fix/commit loop) and `pr-review` (which spawns 5 of each, merges them into one flat pool, and posts a single PR review with inline + global comments). Use this skill when the user asks for "in-depth review", "deep review", "thorough review", "code review without fixing", or invokes
Reviews a pull request with TEN parallel reviewer sub-agents — FIVE `in-depth-review` + FIVE `gh-style-review`, all invoked `--raw`. The orchestrator merges and deduplicates the ten result sets into one pool, keeps findings scoring >= 60, and classifies each as INLINE (specific diff lines) or GLOBAL (broad/architectural). The `gh-style-review` instances also return Discussion Context — prior human comments the diff still leaves open — surfaced as a "Still unaddressed" section. Everything posts as a SINGLE PR review: global findings in full, a names-only list of local findings (also left as inline diff comments), and the unaddressed concerns. Empty sections are never emitted; if nothing survives, NOTHING is posted and the clean result is reported only in chat. A single adversarial reviewer then debates a more-nuanced agent; findings they agree need a change — and that no other reviewer already raised with confidence > 50 — post too, tagged `adversarial`. Use this skill when the user asks to "review this PR", "
Iteratively reviews recent code changes and fixes identified issues or implements improvements. Each iteration spawns SIX concurrent reviewer sub-agents — THREE `in-depth-review` and THREE `gh-style-review` — all invoked with `--raw` against the current branch (PR mode if an open PR exists for the branch, branch mode otherwise). Their raw scored findings are merged and deduplicated across all six instances into one flat pool, filtered to keep anything with confidence >= 50, then fixes are applied and committed one at a time. In PR mode, the gh-style instances also return Discussion Context (which prior human comments the diff resolves vs. still leaves open); the orchestrator surfaces this in the per-iteration summary so the user can see what reviewer feedback is being addressed by the fix loop. The loop stops as soon as one batch finds nothing actionable above the threshold, or after 10 iterations. No GitHub write commands are ever issued. Produces a final summary report. Use this skill when the user asks to
Automatically fix a GitHub PR for the current branch — handles review comments, CI failures, and open questions. Use this skill whenever the user wants to address PR feedback, fix CI, resolve review comments, or generally "clean up" or "fix" their PR. Also trigger when the user says things like "handle PR comments", "fix the build", "address review feedback", "what's failing on my PR", or any variation of wanting to get a PR into mergeable shape. If the user mentions a PR, review comments, or CI checks in the context of fixing things, use this skill.
Opens a new pull request for the current branch with a SIMPLE, scannable title and a short description. If the repo has a PR template (.github/PULL_REQUEST_TEMPLATE.md or similar), fills it in; otherwise uses a default "Goal / how / key-changes-list" format. Designed to avoid the wall-of-text PR descriptions that nobody reads — keep titles single-line, use bullet lists, never enumerate every file in the diff. Use this skill when the user asks to "open a PR", "create a PR", "make a pull request", "submit a PR", "draft a PR", or similar.
Locally mirrors the prompt sent by `@claude review` (the `anthropics/claude-code-action` GitHub Action). Supports two modes: PR mode (arg is a PR number / URL or auto-detected from the current branch) — pre-fetches the same rich PR context the Action injects (formatted_context, PR body, conversation comments, inline review-thread comments, prior submitted reviews, changed files, the diff) and assembles them into the same XML-tagged structure used in `src/create-prompt/index.ts`; branch mode (arg is a git revision range like `origin/main..HEAD`) — degrades gracefully by dropping the comment-derived sections and reviewing just the diff + commit log + changed files. Applies the same review-mode instructions in both modes, then either prints the review to the terminal (when invoked directly) or returns a structured JSON shape (when invoked as a sub-agent by `pr-review` or `review-and-fix`). Output stays local: never posts to GitHub, never opens a PR comment, never updates a Claude comment via MCP. Used as a paral
Reverse Prompt Engineering (RPE) — refine vague or incomplete requests into a precise, LLM-optimized task specification before any implementation begins. Use this skill when a prompt contains `[rpe]`, or when the user explicitly asks to clarify, refine, or spec out a request before acting on it. Do NOT use for requests that are already well-specified and ready to implement.