tdd
Comprehensive guide for Test-Driven Development (TDD) methodology. Use this skill when the user asks to implement features using TDD, write tests first, follow red-green-refactor cycle, or develop code with test-first approach. Also use when user mentions TDD, unit testing workflow, or wants to refactor code with test coverage.
agents-md-review
Review a PR/diff for whether it makes the repo's AGENTS.md or CLAUDE.md stale, and audit those two files for anti-patterns (generic filler, tools named without runnable commands, stale paths). Classifies the change's materiality High/Medium/Low and demands an instructions update for High. Use as the AGENTS.md reviewer in a review panel, or when asked whether a change needs an AGENTS.md/CLAUDE.md update.
babysit-pr
>-
claude-code-session-tracker
Append unfinished Claude Code sessions to a tracker file so they can be resumed via `claude --resume "<session name>"`. Trigger when user signals ending ("done for now", "wrap up", "goodbye"), hits rate/context limit, or asks to save/track the session. Skip short one-offs, clean completions, or trivial Q&A.
codify-writing
>-
create-video-recording
Create PR-ready video verification artifacts for UI, report, browser, visual, or human-reviewable changes. Use when the user asks to record a demo, create a video recording, demo this, make a verification video, produce proof for a PR, prepare the PR Verification section, or when a change is demoable before review. Also use before `gh pr create` for product-visible changes and when building verification reports that should embed or link to a demo video.
eat-like-japanese-local-foodie
Research restaurants the way local Japanese foodies would. Use when Codex needs to find, compare, rank, or document what to eat in Japan for a trip, especially when the user asks for local Japanese recommendations, foodie picks, Tabelog, Hyakumeiten, Michelin Japan, Japanese beef, ramen, coffee, cafes, family-friendly meals, or Japan city/area food surveys.
explain-diff
Use when the user asks for a rich explanation of a code change, diff, branch, or PR — "explain this diff/PR/change", "walk me through what changed", "help me understand this branch". Produces a single self-contained interactive HTML explainer (background → intuition → code walkthrough → quiz). Geoffrey Litt's /explain-diff prompt, verbatim.
goal-audit
Audit a goal against the six qualities of a well-formed goal and rewrite it if it falls short. Triggers whenever the user states a goal, objective, or "what done looks like" for a task — phrases like "the goal is", "I want to", "make this better", "my objective is", "get X working", "build/fix/improve Y" — and whenever the user runs /goal-audit. Use this BEFORE starting work on any non-trivial task so the goal is auditable and has a real finish line, not just when the word "goal" appears. A vague goal gives no completion condition; this skill catches that early.
learn-quiz
Teacher mode after the agent has done work in a session. Verifies the human deeply understands what was done — problem, solution, design decisions, edge cases, broader impact — via incremental restating, gap-filling, and AskUserQuestion quizzes, tracked in a running checklist doc. Use when the user says "teach me this session", "make sure I understand", "quiz me on what you did", "/learn-quiz", or wants to learn from the work just completed. Session should not end until understanding is demonstrated.
live-review
Review an HTML file from Claude with inline comments. Spins up a Claude Code custom channel that serves the HTML with a commenting widget; the reviewer's comments are pushed into a dedicated reviewer Claude session that edits the file and replies in-page. Use when the user produces or has an HTML report/draft/output and wants to comment on it, mark it up, or iterate on it visually — triggers on "review this HTML", "let me comment on", "open for review", "live review", or finishing a draft HTML file. Also triggers whenever the user asks to "create an HTML" (or build/generate/make an HTML file/page/report): after producing it, proactively SUGGEST opening it in live-review — offer the command, don't auto-launch.
loose-ends
End-of-session cleanup sweep. Triggers when the user signals they're wrapping up — "what else to wrap up before I end", "wrap up", "done for now", "anything left", "before I end", "loose ends", "clean up". Sweeps for uncommitted/unpushed work, stale branches & worktrees, leftover files & dev processes, ssh-key state, unverified changes, and unsaved knowledge — auto-doing the safe checks, asking before anything that could lose work.
persona-reader-site
Build a single-file reading site for any person or company from their tweets (ranked by engagement) and long-form writing (blogs, Medium, RSS). Use when the user says "build/refresh a reader site for <person/company>", "scrape <handle>'s tweets and writing into a site", "make Karpathy's writing easier to read", "turn <someone>'s posts into a mini site", or wants a personal index of someone's public output. Backed by Camoufox (stealth Firefox) + a logged-in X profile.
pr-review
Single entry point for reviewing a PR, diff, or branch. Owns the goal-conformance pass — does this achieve its stated purpose, for every input, not just do the lines look fine — and dispatches specialist reviewers (complexity, security, AGENTS.md staleness) in parallel as subagents, then merges everything into one report. Use when reviewing a pull request, diff, or branch, especially when the change instruments, transforms, routes, or tags data over a set of inputs, or claims to do something "for all X".
pr-work-log
Surface off-PR work on the pull request as a maintained comment. Triggers when, while a PR is open, you do work tied to that PR that the diff does NOT show — dispatching/watching GitHub Actions runs, running a backfill/migration/data job, writing to a database, kicking off manual workflows, validation runs, or anything happening in CI/Actions/a DB/an external system. Keeps a single "Off-PR work log" comment up to date as each step lands.
reflection
Interrogate your own just-completed work with three questions — what you're least confident about, what you claimed works but never verified, and the biggest thing you're probably missing. Use when the user runs /reflection, says "reflect on that", "check yourself", or after finishing a non-trivial chunk of work where confident-but-wrong is a real risk.
ship-vercel-app
Deploy a local web app idea from this Mac to production using a new GitHub repository and Vercel GitOps. Use when the user asks to ship, launch, deploy, productionize, or publish a local app/site from ~/code or another local folder, especially with GitHub, Vercel, custom domains, Cloudflare DNS, or "from this Mac mini to production".
ship
>-
understanding-first
>
use-browser
Default protocol for browser automation. Load this skill FIRST whenever any task requires opening a browser, reading a webpage, filling a form, clicking buttons, taking screenshots, scraping data, or any web interaction. Defines the startup sequence, tool priority, and failure protocol so agents always try the right way first and stop on failure instead of guessing alternatives.
write-pr-description
>-
write
>-