@venables/skills

Matt Venables' Claude Code / agent skills, grouped by what they're for.

Reviewing code

Skill	Description
panel-review	Fan a code review out to multiple local CLI agents (codex, claude, opencode) in parallel — each driven through the anyagent CLI — and amalgamate their findings
panel-review-loop	Autonomously loop the review-fix-rereview cycle around panel-review — judge which findings are worth fixing, fix them, re-review, repeat until clean, then report what was fixed, what was left alone (and why), and any goal deviations
branch-review	Review everything on the current branch against origin/main, staged, unstaged, and untracked alike. Findings stay in the session and never reach GitHub
recheck-pr	The second look after a panel-review — adjudicate each prior finding as remediated / explained / moot / outstanding by reading the new code, re-run the panel when the delta is significant, reply to and resolve the threads you're satisfied with, then approve via approve-pr
approve-pr	Approve the GitHub PR you're discussing (usually after a panel-review) — auto-detect the PR from your branch and stamp it with a short fun body (LGTM or a silly emoji) when no message is given, or your message verbatim when one is
pr-review-tab	The tab-lifecycle wrapper for a review spawned by review-prs — run auto-review, close the enclosing Herdr/cmux tab on approval, and with --babysit loop recheck-pr on an interval until the PR is approvable, then close the tab

Posting review findings

Skill	Description
post-panel-review-comments	Triage findings from a panel-review into PR inline comments and Linear tickets via a two-stage select list (auto-falls back to a top-level comment when GitHub rejects an inline)
auto-post-panel-review-comments	Zero-touch twin of post-panel-review-comments — post the legitimate findings straight to the PR (mergeable suggestion when the fix is a drop-in, a +1 reaction when a bot already raised it), routing uncertain / out-of-scope ones to Linear, no prompts
auto-review	One-shot pipeline: run panel-review, post the findings via auto-post-panel-review-comments, then approve via approve-pr with a short LGTM — but only when the PR passes a strict approval gate (full coverage, LOW/polish-only, sound approach, not a draft, and more; see the skill)

Keeping PRs alive

Skill	Description
babysit-pr	Babysit one PR into a mergeable state — resolve its conflicts by merging origin/main, hand its comments to pr-comment-handler, fix failing CI via fix-ci, and push, working in place and auto-detecting the PR from the current branch
babysit-my-prs	Sweep every open PR you authored — run babysit-pr on each that needs it, one dedicated wt worktree per branch (via git-worktree), all in parallel, skipping any PR with no conflicts, no comments, and no failing CI
pr-comment-handler	Read every open review comment on a PR, classify each as fix / push-back / defer-to-Linear, drive the work, then post the right reply on each thread (including Fixed in <sha> replies on push)
fix-ci	Diagnose and fix a failing CI check, on a PR or on main. Distinguishes "this branch broke it" from "main is already red" and fixes the root cause rather than the symptom
sync-main	Bring a long-lived branch up to date with origin/main the same way every time: merge rather than rebase once a PR exists, resolve conflicts, and re-run the checks CI runs before pushing
git-worktree	Create, reuse, or remove a git worktree via the wt CLI so every worktree lands in one consistent place — <repo-parent>/<repo-name>-<branch-slug>, flat alongside the repo. The single source of truth other skills delegate to for "make a worktree"

Building and shipping

Skill	Description
ship	The standing build-to-merged-PR playbook: fresh branch, small commits, a panel review after each phase, sync with main, then a PR with a real body (Problem / Solution, or a Goal for a net-new feature) and Testing
create-pull-request	Open a PR a human can skim: a plain-language title and a body that leads with a short Changes list before the reasoning (Problem / Solution, or a Goal for net-new features) and Testing. Derives both from the diff and commits, honors a repo PR template, and is what ship delegates PR-opening to
tdd-workflow	Enforce test-driven development on new features, bug fixes, and refactors, aiming at 80%+ coverage across unit, integration, and E2E tests
verify	Run the verification loop after a feature or any significant change, checking code quality and conformance before you call it done
startkit	Scaffold a new strict-TypeScript project from the baseline standard — pnpm, native-TS Node, oxlint (type-aware) + oxfmt, vitest, ultra-strict tsconfig, kebab-case filenames, and a GitHub Actions check on PRs and main. Monorepo/CLI/web variants stubbed in references/

Working the Linear backlog

Skill	Description
find-work	Sweep the Linear backlog and act on it — close tickets that already shipped (with evidence and PR links) and pick up small well-specified ones as one branch + reviewed PR each. Grooming is backlog-sweep

Standards and reference

Skill	Description
coding-standards	Universal coding standards and patterns for TypeScript, JavaScript, React, and Node.js
backend-patterns	Backend architecture, API design, database optimization, and server-side practice for Node.js, Express, and Next.js API routes
frontend-patterns	Frontend patterns for React and Next.js: state management, performance, and UI practice
security-review	The checklist and patterns to reach for when adding auth, handling user input, working with secrets, creating API endpoints, or touching payment paths
find-docs	Fetch current documentation, API references, and examples for any library, framework, SDK, or CLI tool, rather than trusting training data that may predate the latest release
optimize-agents-md	Create, audit, or prune AGENTS.md / CLAUDE.md so every AI coding tool reads one lean source of truth. Handles monorepos, too

Utilities

Skill	Description
bro	Restate the last message in plain human language, no jargon — one human talking to another. User-invoked only. Credit: Dillon Mulroy

License

MIT