MasterClaude

Scope & verification discipline for a working session. Triggers when about to claim "done / fixed / passing", when a test is about to be weakened, skipped, commented out or deleted to get a suite green, when a refactor, abstraction or dependency appears that the task never asked for, and on "audit this diff" / "did we actually verify that?". Guidance the model follows — it is not enforcement; nothing here can block a tool call.

The MASTER CLAUDE leader/conductor. Triggers on "master claude", "set up master claude", "mc setup", "onboard me", "build my team", "what's new", or when starting work in a new/unfamiliar project. It interviews you (grill-me), maps the project, then assembles a tailored team from the installed skills and agents — Sentinel the project cartographer plus planning, review, understanding, guardrail and security specialists — and runs it on your work. It builds the best `.claude/` workspace per project, delegates to parallel subagents (choosing the right model per agent), brainstorms hard and decides fast, keeps a complete view of every installed capability, keeps itself and you current with the newest Claude Code features, and proactively offers the right tool the moment a need shows up.

Continuity across long sessions. Snapshots in-progress work before context compaction and restores it after, and nudges auto-compaction toward a natural breakpoint so nothing important is lost mid-task.

Stay current with Claude Code and Claude itself. Triggers on "what's new", "what can Claude Code do now", "is there a newer/better way", "update master claude", "which model should I use", or before relying on any claim about a Claude Code feature, model, price, or limit. Checks the installed version, the live changelog, and the harness surface MASTER CLAUDE depends on — then reports only the deltas that matter for this project.

How MASTER CLAUDE picks the right model for each agent and task. Triggers on "which model", "pick a model", "model selection", "use haiku/sonnet/opus", "cheaper model", "model per agent", or whenever spawning a subagent/team and choosing its model, or optimizing cost vs. quality on a long run. Gives a role→model heuristic, the "turn count beats token price" rule, the effort dial, and the cost math so the team is both strong where it matters and cheap where it doesn't.

How MASTER CLAUDE delegates work to subagents and teams. Triggers on "delegate", "spawn agents", "parallel agents", "fan out", "run these in parallel", "subagent", "orchestrate", "build a team for this", or whenever a task splits into independent chunks, needs a fresh-eyes review, or means reading many files. Teaches the orchestrator-worker pattern: pick the right primitive (subagent vs agent-team vs background session), write a complete delegation brief, parallelize only when it pays, scale effort to the task, and verify with an adversarial reviewer. Use to plan and run any multi-agent push.

Run parallel or risky work without collisions using git worktrees. Triggers on "worktree", "isolate this change", "try this without touching main", "run these agents in parallel safely", "spike/experiment", or whenever multiple subagents or features would edit the same tree at once. Each task gets its own branch + working directory, so concurrent work never clobbers another's files and a failed experiment is thrown away by deleting a folder. Pairs with subagent-orchestration for true parallel building.

Turn an approved spec + decomposition into a precise, zero-context implementation plan document that another engineer (or a fresh subagent) can execute with no prior knowledge. Triggers on "write the plan", "write an implementation plan", "turn this spec into a plan", or after grill-me/cap-spec-smith + cap-decomposer have produced a spec and a step list, before any code. Each task carries exact file paths, interface contracts, real test code, exact commands and their expected output — no placeholders. The plan is the contract; cap-execute-plan (or parallel subagents) runs it.

Before you push, it reviews your diff as an unsentimental reviewer of your own work: issues labeled blocker/major/minor/nit with file:line and a concrete fix, plus missing tests and edge cases, ending with a verdict and the one thing to fix first.

The git mechanics MASTER CLAUDE runs by default — invoked by the leader, not auto-triggered. Covers session awareness (branch, dirty count, ahead/behind), branch-before-you-build (mc/<slug>, never the default branch), commit at every green step (conventional, Fixes #N), a WIP commit before risky bash, finishing a branch (PR/merge/delete), worktrees for parallel writers, and the bright line: local git aggressive by default, remote git always consented, never auto-push. The off switch lives in .claude/master-claude.json.

Drive an approved plan or spec to done — supervised, one verified step at a time. Triggers on "execute the plan", "work the backlog", "build it from the plan/spec", "implement this plan", or right after cap-plan-first / cap-decomposer / grill-me produced a plan. Works the tasks in order, verifies each (build/tests) before moving on, checkpoints with you at the risky/irreversible steps, and keeps a progress trail. You stay in the loop the whole way, verifying each step.

Know which environment you are standing in before you touch anything. Triggers on starting work on a server, "is this prod?", "which environment is this", before any deploy, migration, restart or seed, and the moment you ssh into a box you didn't just provision. Writes the verdict to `.mc/env` so every other ops skill can key off it.

Incident commander for a production emergency. Triggers on "prod is down", "we're seeing 500s", users who can't log in, check out, or load the page, an alert firing, a pager going off, or "the site is slow" in a production context. Stabilizes first, diagnoses second, and keeps a timeline — the advisor, never the one typing on the prod box.

Read the system's own story before you touch it — logs, metrics, health, recent deploys, error rates. Triggers on "why is it slow", "what's happening on the server", "check the logs", "is it healthy", and as the first step of `ops-incident`. Read-only toward the system: it gathers evidence and files findings, it never mutates the box.

The blameless post-mortem that closes the loop after an incident. Triggers on "post-mortem", "how did this happen", "how do we prevent this", a retro after any outage, or a failure that has now recurred. Turns a timeline into contributing causes, action items with owners, and one concrete artifact that would have caught it — built, not admired.

Revert a bad deploy, by deploy shape. Triggers on "roll back", "revert the deploy", "the deploy broke it", "undo the release", or right after `ops-incident` picks rollback as the stabilizing lever. Covers the shape you actually shipped — git, tag, image, or flag — and the one-way door that is a schema migration.

The pre-deploy gate. Triggers on "deploy", "ship it", "release", "cut a release", "push to prod", "roll it out", or editing a CI or deploy config. Produces a deploy brief with five required slots and a tagged release — or it refuses to call the deploy ready.

Architecture Decision Records — capture a hard-to-reverse choice while the reasons are fresh. Triggers on "why did we choose", "record this decision", "ADR", picking a database, framework, protocol, auth model, or API shape, or any fork where reversing the choice later would cost more than a day. One short file per decision, so the next person inherits the reasoning, not just the result.

Turn the team's findings into GitHub issues, and keep them in sync. Triggers on "file these as issues", "sync findings", "open issues for the security review / the QA run / Sentinel", "close the fixed ones", or after a sweep/audit that produced findings. Local finding files stay the source of truth; issues are the human surface. Security findings on a public repo are held back by default.

Design intelligence — the concrete picks, not just the taste. Queries a local database of product types, color palettes, UI styles, font pairings and UX rules, and returns an actual design system: real hex tokens, a named style, a font pairing, decision rules and the anti-patterns to avoid. Triggers at the START of any UI work — a new page, app, landing page, redesign, component library — and on "what colors / font / style should this use", "make it look designed, not generic", or "it looks AI-generated". Persists the result to .mc/design/MASTER.md so every later session builds the same product.

The guardrail switchboard. Triggers on "which guardrails are on?", "turn on / turn off <guardrail>", "set up guardrails for this project", or when a guardrail keeps firing and the developer wants it quieter. Records the choice in .claude/master-claude.json so it survives the session; the leader honours it.

Dependency supply-chain check before a package lands. Triggers when adding or bumping any dependency — package.json, requirements.txt, pyproject.toml, go.mod, Gemfile, Cargo.toml — or when running npm/pnpm/yarn/pip install, and on "is this package safe / real / maintained?". Catches hallucinated and typosquatted names, known-vulnerable versions, and weak trust signals. Guidance the model follows before it writes the dependency — it is not interception.

Generate mc.html — a self-contained, CLI-styled, professional, charted dashboard at the project root showing MASTER CLAUDE's team and its findings for THIS project (test results, coverage, audit/map findings), led by MASTER CLAUDE. Triggers on "make the dashboard", "mc.html", "show the project status", "generate the team report", or as the final step of wf-tester. A local artifact — never published to any website.

Accessibility done properly — WCAG 2.2 AA, keyboard, screen readers, focus, contrast. Triggers on "is this accessible", "a11y / WCAG / ADA / Section 508 / EAA compliance", an accessibility audit or complaint, a failed axe/Lighthouse a11y score, or any custom interactive control (modal, menu, combobox, tabs, drag-drop, carousel). Also the first gate before shipping any UI. Returns findings mapped to a WCAG criterion with the concrete fix.

Review UI before you ship it, like a picky designer. Triggers on "review this UI / page / component", "does this look good / professional", "is this accessible", before merging frontend work, or after building any interface. Returns findings labeled blocker / major / minor with the fix — across visual hierarchy, spacing, contrast & a11y, responsive behavior, interaction states, and consistency with the design system.

Set the design foundation before building any UI, so the output looks designed, not default. Triggers when starting a new frontend, a redesign, a landing page, or a component library, and on "make it look good/professional", "it looks generic / like Bootstrap / AI-generated", or any UI work with no established visual system. Establish or detect design tokens (color, type scale, spacing, radius, shadow, motion), a font pairing, and a component vocabulary — reusing shadcn/Radix/Tailwind conventions — before writing components.

Motion that means something — transitions, micro-interactions, page and list animation. Triggers on "add animation / transitions / make it feel smooth", "it feels janky / cheap / static", a hero or scroll effect, a modal/drawer/accordion/toast entrance, list reordering, or any Framer Motion / GSAP / View Transitions work. Covers duration, easing, what to animate, reduced-motion, and the line between polish and nausea.

Frontend performance where users feel it — Core Web Vitals, bundle size, render cost. Triggers on "it's slow / laggy / janky", "the page jumps", a poor Lighthouse or PageSpeed score, LCP/CLS/INP problems, a large bundle, slow lists or re-renders, or before launching anything public. Diagnoses with real numbers first, then fixes the thing that actually costs — not the thing that's fun to optimize.

Decide what to build before deciding how it looks. Triggers on a new product, feature or redesign, on "who is this for", "what should this screen do", "users are confused / churning / not converting", on vague briefs ("make a dashboard"), and before any information-architecture or navigation decision. Produces the user, the job-to-be-done, the primary flow, and the one thing each screen must accomplish — the brief the rest of the design team builds against.

Run a full design pass over a product — establish who it's for, generate and persist a real design system, rebuild the UI against it, then review front-to-back and report what changed. Use when a whole frontend needs to go from generic to designed: a redesign, a pre-launch polish, "make this look professional", "our UI looks AI-generated", or a new product with no visual system. For one page or one component, use the focused fe-* skill instead.

Run the team across SEPARATE, parallel Claude Code sessions for real throughput — beyond in-session subagents. Triggers on "make it faster", "too slow", "parallelize", "run the team in separate sessions", "fan out", "background agents", "agent teams", "run several at once", or a big batch of independent work. The leader decomposes the job into independent chunks, dispatches each to its own session (true OS-level parallelism), monitors, and integrates — with hard guardrails on cost and isolation, because N parallel sessions burn usage N× and editing the same files in parallel causes merge hell.

Test a deployed surface from the outside as a black box — no source, no test hooks: functional smoke, visual regression, accessibility (axe), performance (Lighthouse), SEO, PWA (manifest/service worker/offline), broken links, and console/network errors. Triggers on "black-box test", "test the live site", "check the deployed app", "lighthouse/a11y/seo audit of the site". Not for unit tests.

Test at the code level — unit and integration tests, coverage of the critical paths, and test-first (TDD) discipline for new work — reusing the project's own test runner. Triggers on "add tests", "increase coverage", "unit/integration tests", "TDD", "the code has no tests", "test this function/module". Not for outside-in UI checks (use test-blackbox) or load (use test-stress).

Load, stress, spike, and soak testing — find where a service bends or breaks under pressure and where latency/error rates cross the line. Uses k6 (scriptable, low-overhead). Triggers on "load test", "stress test", "how much traffic can it handle", "test under pressure", "spike/soak test", "will it hold up at launch". Not for single-user functional checks (use test-blackbox/test-user-end).

Test a project the way a real end user experiences it — click-through journeys, forms and validation, empty/error/loading states, i18n/RTL, keyboard, offline/PWA — using the QA engine's agent and human lanes. Triggers on "test the user flows", "does the UI actually work", "end-user testing", "test signup/checkout/onboarding as a user". Not for code-level unit tests (use test-code).

Drive a full QA pass on any project as a MASTER CLAUDE-led tester team — user-end, black-box, code, and stress/load — then emit a charted mc.html status dashboard. Triggers on "test everything", "QA this", "full test pass", "is it ready to ship", "pre-release check", "the app has no tests", "test this app end to end". Not for a single unit test (use cap-tdd) or a security audit (use wf-security-audit).

Curate the model's context window for the most signal per token — stop burning tokens on bloat. Triggers on "context engineering", "burning tokens", "context window full", "reduce token usage", "the model lost the thread", "too much in context", "prompt caching", "session handoff", or any long / expensive session. The discipline of deciding what goes in the window (and what doesn't): cache-stable instructions, retrieve don't dump, fresh docs at query time, audit your MCPs, compress tool output, and measure tokens like an infra cost.

Detect and de-identify PII before sensitive data reaches an AI model, a log, or an external service. Triggers when an AI pipeline sends user data to an LLM, when handling names / emails / phone / national-id / credit card / IBAN / medical records / addresses, and on "anonymize", "redact", "mask PII", "sanitize before sending to the model", "GDPR / HIPAA", or "don't leak user data to the AI". Builds a redaction step with Microsoft Presidio (analyze → anonymize) — masking, redacting, hashing or encrypting PII, with built-in or custom recognizers, for text, images and DICOM.

Build UI components that are accessible, responsive, and complete — not just the happy path. Triggers when building a button, form, input, modal, dropdown/menu, table, card, nav, toast, tabs, tooltip, or any reusable component — or when UI "works but feels off", "isn't accessible", or "breaks on mobile". Covers a11y (semantics, ARIA, keyboard, focus), every interaction state, responsive behavior, and the polish that separates designed from default.

Turn a screenshot, a brand, a competitor, or a vibe into real UI — and pick the right Claude output channel for it. Triggers on "build this <screenshot>", "make it look like <site>", "match our brand", "design a hero / poster / social card", "a chart / diagram / mockup", "generate a UI for…", or design-to-code. Maps the request to a Claude artifact, the visualize widget, or the canvas-design / theme-factory / algorithmic-art skills, then reproduces the style faithfully (original assets only).