Claude Code Platform Reference

Skills (ls .claude/skills/ for the live count)

Orchestration: /planner, /builder, /cycle
Engine: /rust-engine, /build, /arch-validator, /rust-best-practices
Web: /frontend, /mcp-commands, /web-accessibility, /shadcn, /vercel-react-best-practices, /vercel-composition-patterns, /vercel-react-view-transitions
Next.js: /next-best-practices, /next-cache-components, /next-upgrade
Testing: /testing, /playwright-best-practices, /tdd
Infrastructure: /infra-services, /troubleshoot, /kanban, /babysit-prs, /pr-code-review, /pr-green-machine, /env-health-check, /changelog-review, /deploy-to-vercel, /resolve-pr-comments, /resolve-all-pr-comments, /api-middleware-migrate, /autonomous-sprint
Database: /db-migrate, /neon-postgres, /claimable-postgres, /neon-postgres-egress-optimizer
Deployment: /deploy-engine (user-only)
Billing: /stripe-webhooks (background, auto-loads on billing edits)
Features: /game-engine, /game-ui-design, /multiplayer-readiness, /viewport
Workflow: /design, /architect-flow, /docs, /developer-experience
Review: /review-protocol, /component-checklist
Vercel: /vercel-cli-with-tokens

MCP Servers (.mcp.json — 8 servers)

• context7 — live library documentation for all 30+ dependencies
• neon — direct Neon Postgres queries (needs NEON_API_KEY)
• playwright — browser automation for E2E verification
• github — GitHub API access for PR/issue/Actions operations
• sentry — error tracking, issue search, event analysis
• stripe — payment processing, webhook debugging, test cards
• upstash — Redis cache operations and rate limiting
• taskboard — the local ticket board (launched through a git alias; see docs/guides/taskboard-sync.md)

Claude Code defers MCP tools behind Tool Search by default. Only context7 sets
"alwaysLoad": true, so its tools are in context from session start: it needs no
credential, only reads documentation, and nearly every session uses it. Every other
server stays deferred on purpose. neon, sentry, stripe and upstash read
credentials from the environment, and github needs a GitHub token and can write to
the repository (push, merge, create and delete files), so loading its tools into every
context, reviewer seats included, is not worth the saved search. alwaysLoad also
makes startup wait for the server. Add it to another server only if that server needs
no credential, cannot write, and is needed in nearly every session. The key and its behavior are documented at
https://code.claude.com/docs/en/mcp#exempt-a-server-from-deferral. The key affects
Claude Code only: Codex does not read .mcp.json.

Hooks (.claude/hooks/ — 52 scripts, 20 event types)

Event types: PreToolUse, PostToolUse, SessionStart, SessionEnd, UserPromptSubmit, Stop, SubagentStart, SubagentStop, TaskCreated, TaskCompleted, TeammateIdle, PreCompact, PostCompact, StopFailure, ConfigChange, InstructionsLoaded, WorktreeCreate, CwdChanged, PostToolUseFailure, FileChanged

Hook	Event	Purpose
inject-lessons-learned.sh	PreToolUse (Edit/Write/Bash)	Anti-patterns before action
pre-push-quality-gate.sh	PreToolUse (git push)	Lint+tsc before push
block-main-commits.sh	PreToolUse (git commit)	Prevents commits on main
verify-branch.sh	PreToolUse (Edit/Write)	Prevents edits on wrong branch
check-db-transaction.sh	PreToolUse (Edit/Write)	Warns on db.transaction()
review-quality-gate.sh	Stop (reviewers)	Validates PASS/FAIL verdict
block-writes.sh	PreToolUse (reviewers)	Reviewers are read-only
builder-quality-gate.sh	Stop (builder)	Verifies lint+tests ran
cargo-check-wasm.sh	PostToolUse (rust-engine)	Cargo check after .rs edits
validate-agent-output.sh	SubagentStop	Verify reviewer output
reject-incomplete-review.sh	TeammateIdle	Exit 2 if no verdict
save-critical-context.sh	PreCompact	Dump decisions before compaction
restore-context-hints.sh	PostCompact	Inject context after compaction
on-session-start.sh	SessionStart	Taskboard state + sync
on-stop.sh	Stop / SessionEnd	Worktree safety commit + GitHub sync

Validation Tools (.claude/tools/)

• validate-rust.sh — Architecture boundaries, bridge isolation, unsafe audit
• validate-frontend.sh — ESLint, TypeScript (web, packages/ui, apps/docs), vitest
• validate-mcp.sh — Manifest sync, MCP type-check + tests, AI parity audit
• validate-tests.sh — Test inventory, coverage report
• validate-docs.sh — Documentation integrity, version refs
• dx-audit.sh — Cross-IDE consistency, tool health, onboarding check
• validate-all.sh — Run all validators

Plugin (.claude-plugin/plugin.json)

Internal plugin manifest. Test with claude --plugin-dir .

Approved Specs

• specs/2026-03-25-game-creation-orchestrator-phase2a-v4.md — Game Creation Orchestrator (systems-not-genres)