From Agents Dev Kit
Review, audit, sanity-check, look-at any review-able target. Triggers on a GitHub PR URL (fetched via the gh CLI), a local path or "." (review the working tree), a markdown/doc file, or a comment-thread URL. Read-only by default; --fix applies accepted findings locally and pushes to the PR branch after confirmation (never force, never merge, never to a protected branch). Produces severity-tiered findings (blocker / critical / should / may / nit) with path:line and <=15-word verbatim evidence quotes. Six dimensions: correctness, tests, security, performance, readability, consistency. For a deep PR review with cross-file code-context retrieval, use /adk:pr-review.
How this skill is triggered — by the user, by Claude, or both
Slash command
/adk:review <target> [-i|--interactive] [--fix] [--severity blocker|critical|should] [--dimensions a,b,c] [--deep]<target> [-i|--interactive] [--fix] [--severity blocker|critical|should] [--dimensions a,b,c] [--deep]This skill is limited to the following tools:
The summary Claude sees in its skill listing — used to decide when to auto-load this skill
Polymorphic on the target. **Read-only by default.** Lightweight by design — no worktree, no embeddings. For a deep PR review with code-context retrieval and inline comment posting, use `/adk:pr-review`.
Polymorphic on the target. Read-only by default. Lightweight by design — no worktree, no embeddings. For a deep PR review with code-context retrieval and inline comment posting, use /adk:pr-review.
The full operating contract lives in this skill folder — read these as you need them:
| Aspect | File |
|---|---|
| How you review (voice, tiers, evidence rules) | persona.md |
| The phased process + Workflow orchestration | workflow.md |
| Hard rules + refusals + safety | rules.md |
| Target routing (PR / local / doc / thread) | dispatch.md |
dispatch.md and classify the target. Resolve the diff:
gh pr view <url> --json … for metadata, gh pr diff <url> for the diff. (GitHub access is always the gh CLI — assume the user ran gh auth login.). or local path → git diff against the merge-base of the current branch.persona.md — adopt the findings-first reviewer stance.workflow.md. For any non-trivial diff, fan the six dimensions out in parallel with the Workflow tool (one code-reviewer / security-auditor agent per dimension), then adversarially verify each finding before it survives. Small diffs (≤ ~150 LOC, single concern) may be reviewed inline without a workflow — but still one dimension at a time.path:line + ≤15-word quotes, then a one-line ship | iterate | reject recommendation. Posting to a PR or applying --fix is gated per rules.md."Always have a workflow." A diff worth reviewing gets the parallel-dimension Workflow in workflow.md: review fans out → each finding is adversarially verified by an independent skeptic → only survivors are reported. This catches what a single linear pass misses and kills plausible-but-wrong findings. Skip the Workflow only for a trivial diff, and say so.
-i — walk each finding with the user (accept / reject / edit) before it lands in the report or gets posted.--fix — after the review, apply accepted findings locally, validate, and push to the PR's head branch after explicit confirmation. Never force-push, never merge, never push to a protected branch (rules.md).--deep — use a stronger reasoning profile; auto-select for security-sensitive, cross-module, or ambiguous diffs.npx claudepluginhub sujeet-pro/agents-devkit --plugin adkGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.