skills

Continue an in-progress /grill-me session in auto-pilot — the agent keeps asking the same relentless questions, but answers each one with its own best recommendation and moves on. The user no longer types answers; they interrupt (Esc) only when they disagree. Use to power through long grilling sessions once you and the agent are clearly aligned.

Hand a goal-maker board off to Codex (gpt-5.5, low reasoning by default) — Claude validates the board, runs `codex exec` with the /goal directive, captures the output, re-validates, and reports state changes. Use after /goal-maker bootstraps the charter and state.yaml when you want Codex (cheaper at low reasoning, separate context) to actually execute the PM loop while Claude Code stays the orchestrator. Triggers on "codex-goal", "run goal in codex", "delegate goal to codex", "hand to codex", or "/goal" when the user is in Claude Code (which doesn't have /goal natively).

Cross-agent code review: run OpenAI Codex to review your changes, then address its feedback. Use when asked to "codex review", "get a second opinion", "cross-review", "have codex review this", or "external review". Works with uncommitted changes, specific commits, or branch diffs.

Set up a rolling Scout/Judge/Worker task board with a charter, machine-truth state.yaml, and durable receipts for long-running autonomous coding work. Use when a goal is broad, multi-hour, ambiguous, recovery-flavored, or needs serial discipline (one active task at a time) instead of parallel teammates. Triggers on "goal-maker", "set up a goal", "create a charter", "rolling task board", "scout/judge/worker", "PM-owned board", or when the user wants long-running autonomous work that's NOT a parallel swarm and NOT a multi-milestone mission. Pairs with /grill-me upfront and /principles

Measure and lower CRAP (Change Risk Anti-Patterns) — a metric that fuses cyclomatic complexity and test coverage to surface fragile code. Use when verifying test quality after a feature, gating PRs, refactoring legacy code, setting up a new project's quality bar, or whenever the user mentions "CRAP score", "complexity", "Change Risk Anti-Patterns", or wants to know which functions are most likely to break. Pair with mutation-testing so coverage is meaningful, not superficial.

Bootstrap CLAUDE.md as a short table-of-contents plus a structured docs/ directory (agent guidance, architecture, product specs, acceptance tests, ADRs, exec plans, quality grades). Use when CLAUDE.md is missing, when asked to "create CLAUDE.md", "bootstrap project", or "set up agent context".

Invoke this skill anytime you are creating a plan—such as crafting an RFC, writing a PRD, starting a new app, or proposing an architecture change. It will relentlessly interview the user to resolve all dependencies and edge cases.

Multi-milestone autonomous development for large goals. Decomposes a big objective into milestones with validation gates. Orchestrator plans, workers build via TDD in isolated worktrees, independent validators exercise the system as a black box against a validation contract. Fix features close gaps. Converges until all milestones pass. Use when the user says "mission", "missions", wants to build something large (multi-feature, multi-day), or needs structured autonomous development beyond a single backlog.

Run and interpret Stryker mutation testing; kill survivors to reach ≥95% score. Use when running mutation tests, setting up Stryker, interpreting survivors, or verifying test quality after TDD.

Enforce the "parse, don't validate" discipline at every system edge. Data crossing a trust boundary must be parsed through a schema before entering application logic. Language-agnostic — covers TypeScript, Python, Go, Swift, and Kotlin. Use when writing code that receives external data (API responses, HTTP requests, env vars, file reads, message queues) or when reviewing code that casts, asserts, or accesses unvalidated external input.

The 10 operating principles for building software with AI agents — when to use them, what failure mode each one prevents, and which skills operationalize them. Use when bootstrapping a new project, reviewing an agent workflow that drifted, deciding between tactical fixes and structural changes, or whenever the conversation turns from "what to build" to "how to build it well". Triggers on "principles", "commandments", "how should we work with agents", "what's our philosophy", or when an agent loop is producing low-quality output and you need to diagnose which principle was violated.

A multi-agent AI development workflow. Front-loads planning questions (grill-me), then launches parallel agents in Git worktrees: a Feature agent builds from a backlog via TDD, while CRAP, Mutation, and Acceptance agents continuously harden the codebase using deterministic metrics. Use when the user says "swarm", "run the swarm", "kick off a swarm", or wants to plan and then execute a long-running multi-agent run.

A hypothesis-generating loop that validates itself, run entirely inside Claude Code. Claude proposes a change, runs a fixed-budget experiment, measures a deterministic metric, then git-commits wins and git-reverts losses — repeating ~10–100x while you're AFK. The harness is `/loop` (single agent iterating) or `/agent-teams` (a proposer + a validator/judge running in parallel) — no external scripts, full progress visible in the session. Use whenever there's a single scalar metric to push (lower KL / RMSE / loss / latency, higher accuracy / score / win-rate / pass-rate) and the validation runs end-to-end without human judgment. Works on simulators, fitting routines, portfolio optimizers, eval pass rates, hyperparameter tuning, solver tuning, fuzzing, perf budgets, prompt scores — anything optimizable. Triggers: "autoresearch", "run autoresearch", "optimize this overnight", "AFK optimize", "autonomous experiments", "iterate until <metric> hits <target>".

Spawn and coordinate Claude Code Agent Teams — multiple Claude sessions sharing a task list and mailbox, with one lead coordinating teammates that can message each other directly. Use when the user wants parallel research, competing-hypothesis debugging, multi-perspective code review, or cross-layer feature work that's too coordinated for /nightshift or /swarm but too multi-agent for plain subagents. Triggers on "agent team", "claude-teams", "spawn teammates", "parallel review", "competing hypotheses", "team of claudes", or when the user mentions cmux claude-teams or CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS.

Manage cmux terminal workspaces and browser surfaces for parallel AI agent sessions. Create, switch, monitor, and communicate between named workspaces. Automate embedded browsers with navigation, DOM interaction, inspection, console/error capture, and session state management. Use when running multiple Claude Code agents, when /swarm or /nightshift need terminal-level orchestration, when automating browser testing via cmux surfaces, or when the user mentions "cmux", "workspaces", "parallel agents", "orchestrator", or "cmux browser".

Create AFK Ralph loop scripts for any task. Interrogates the user to nail down requirements, then generates a complete .ralph/ directory with loop script, backlog, progress tracking, and lessons file. Use when the user wants to run an autonomous Claude loop, create a ralph script, go AFK on a task, run something in a loop while away, batch-process work autonomously, or mentions "ralph", "ralph creator", "afk loop", "create a loop", or "run this while I'm away".

After finding a bug caused by insufficient test coverage, brainstorm test cases across the entire feature area that could uncover similar bugs. Systematically explores edge cases, boundary conditions, state transitions, and integration points to identify coverage gaps before they become production incidents.

Enforce a hard-cut cleanup policy: keep one canonical implementation and delete compatibility, migration, fallback, adapter, coercion, and dual-shape code. Use for pre-release or internal-draft refactors where the goal is one final shape, especially when changing schemas, contracts, persisted state, routing, configuration, feature flags, enum/value sets, or architecture.

Two-phase design loop like /swarm but for visual design. EXPLORE: launches N parallel subagents in Figma — each designs a distinct concept on its own page simultaneously, then cross-model consensus (Claude + Codex) picks the winner. DEEP: iterates the chosen concept with critique → impeccable skills → fixes until it converges. Use when starting a new UI from scratch or iterating an existing design, when asked to "autodesign", "explore designs", "design this in Figma", "go wide then deep", or "make it look good".

Autonomous AFK development loop. Sequentially works through a specs/bugs backlog: grill-me preflight → evals-first surface → pick task → testing-trophy TDD → progressive commits → eval stack (hooks → integration → e2e + screenshots → LLM judges → codex review) → commit → morning briefing. Use when going AFK — lunch, gym, overnight, weekend. Optionally provide duration so the agent can scope work accordingly.

Test actor-kit state machines in Storybook using mock clients and play functions. Covers static snapshots, interactive state transitions, event interception, and multi-actor nesting. Especially suited for game UIs with complex state progressions. Use when building Storybook stories for actor-kit components, testing game states, or setting up play function interactions with mock actor clients.

Integrate actor-kit with TanStack Start/Router for server-rendered, real-time stateful apps on Cloudflare Workers. Covers route loaders, server functions, SSR hydration, WebSocket handoff, middleware, and E2E testing with Playwright. Use when building a TanStack Start app with actor-kit, setting up actor-kit server functions, or configuring the SSR-to-WebSocket hydration flow.

Create and maintain Architectural Decision Records (ADRs) with date-named files sorted like migrations. Use when making structural decisions (new deps, pattern changes, tech choices), when asked to "record a decision", "create an ADR", "log an architecture choice", or "why did we do it this way".

Monitor a PR through CI, diagnose and fix test failures, resolve merge conflicts, post QR codes for mobile preview builds, and auto-merge when ready. Use when asked to "babysit", "monitor this PR", "watch CI", "fix CI", "post QR code", "make sure CI passes", or "merge when green".

Automated QA for WebGL/Three.js/R3F multiplayer games built with actor-kit. Playwright visual regression, game state assertions, test harness setup, structured logging, and bug-finding loops. Use when asked to "QA the game", "test the game visually", "find bugs", "set up game testing", "visual regression", "playwright for webgl", "game test harness", or "game logging".

Minimize unnecessary React re-renders when consuming external state (XState, @xstate/store, Zustand, Redux, Nanostores, context). Prefer selector-based subscriptions over useState(wholeObject). Use when dealing with external state in React, optimizing re-renders, choosing state patterns, or integrating with these libraries.

Focuses exclusively on writing robust integration tests at system boundaries (seams) rather than writing brittle, shallow unit tests. Use when adding test coverage to an existing system, or testing an integration between two distinct modules.

Storybook play function integration testing for TanStack Start/Router apps. Covers Vite config for blocking server code, router/query/theme decorators, state mocking strategies (props, collections, React Query cache, actor-kit), play function patterns, viewport coverage, and vitest browser test runner. Use when writing Storybook stories, setting up Storybook in a TanStack Start project, testing route-level pages, or adding play function interaction tests.

Kent C. Dodds' Testing Trophy: write more integration tests, fewer unit tests, confidence over coverage. Covers React (Storybook + play functions), Cloudflare Workers (vitest-pool-workers + D1), and Swift (XCTest UI + snapshot). Use when deciding test strategy, writing new tests, or reviewing test quality.

Formal verification of system designs using TLA+ and the TLC model checker. Models concurrent state machines, finds race conditions, deadlocks, and invariant violations before any code is written. Also verifies pure algorithm invariants by exhaustively checking all input combinations. Use when the user says "tlaplus", "tla+", "formal verification", "model check", "verify my design", "check for race conditions", "state space", "verify concurrency", "find bugs", "verify algorithm", or when working on systems with shared mutable state, offline sync, queues, distributed coordination, multi-agent orchestration, or complex pure functions with subtle invariants.

Verified Spec-Driven Development (VSDD) — a rigorous AI-native engineering methodology that fuses Spec-Driven Development, Test-Driven Development, and adversarial verification into a single pipeline. Use when the user says "vsdd", "spec-driven", "verified spec", "adversarial review", "zero-slop", "full verification", "spec-first development", "build this properly with specs", or wants a rigorous spec → test → implement → verify workflow. Also use when asked for formal verification, mutation testing workflows, or adversarial code review as part of a development process.

Write integration tests for Cloudflare Workers using vitest-pool-workers and cloudflare:test. Tests the full HTTP cycle — request through handler, middleware, bindings (D1, KV, R2, DO, Hyperdrive), and back. Use when adding test coverage to a Worker, when a new route/endpoint is created, or when /nightshift, /swarm, or /ralph-tdd encounter a Cloudflare Workers service.

Set up and use the `emulate` package for local API emulation in tests and CI. Covers CLI usage, programmatic API (`createEmulator`), seed config, and integration test patterns for GitHub, Vercel, Google, Slack, Apple, Microsoft, and AWS emulators. Use when configuring emulate, writing tests against emulated APIs, or setting up CI pipelines with emulated services.

Sharpen product design judgment for software UI/UX, interaction flows, jobs-to-be-done, hierarchy, trust, governance surfacing, and competitor-informed critique. Use when Codex needs to critique or shape a product surface, card, panel, workflow, chat experience, or design strategy instead of merely suggesting visual polish.

Verify UI changes visually using browser automation before telling the user it's done. Use when making CSS, layout, or visual changes to web apps — especially Starlight docs sites, landing pages, or any frontend work. Triggers on "does this look right", "verify the UI", "check the layout", or after any CSS/HTML change.

Evals-first development: write evaluations before specs, code, or designs. Evals become the guardrails that condition the agent to write better specs, clearer requirements, and code that works on the first try. Use when starting a new feature, project, or initiative. Use when the user says "evals first", "start with evals", "write evals", "evaluation-driven", or wants to establish quality guardrails before building. Also use when converting guides, papers, or domain knowledge into enforceable evals.

Pull all Jira tickets assigned to me that aren't done and help create copy-pasteable prompts to kick them off in other Claude Code or Cursor sessions. Use when asked to "plan my tasks", "what should I work on", "task planner", "generate prompts for my tickets", or "kick off my Jira tickets".

Structured debugging for production and staging issues. Maps symptoms to tools, queries, and data sources (Sentry, PostHog, wrangler logs, console, simulator logs). Use when something is "not working", "broken", "failing in prod", "500 error", "check logs", "check sentry", "why is this happening", or when debugging a deployed service or mobile app.

Deploy Cloudflare Workers and verify changes work in staging/preview. Use when asked to "deploy", "ship", "push to staging", "deploy and test", "verify deploy", or "check staging".

Build, install, and test Expo/React Native apps on simulators and physical devices. Use when asked to "run on simulator", "install on device", "test on phone", "run detox", "preview build", or "build and test".