Skills (Claude Code 2.1+)

In addition to commands, the project includes 53 Skills in .claude/skills/:

Trigger column. A phrase list means the skill auto-loads when the conversation matches it. manual only means the opposite: the skill carries disable-model-invocation: true, so the model cannot load it — not on a phrase match, and not from inside the same-named command either. Only you can start it, by running the slash command shown. Pinned by tests/skills-catalog-drift.bats, so a row and its frontmatter cannot drift apart.

Core skills

Skill	Trigger	Context
dev-tdd	"TDD", "test first", "write the tests"	inline
work-commit	manual only — run /work-commit	inline
dev-debug	"bug", "error", "debug"	inline
qa-review	"review", "code review"	fork
qa-security	"security audit", "OWASP"	fork
work-plan	manual only — run /work-plan	inline
work-explore	manual only — run /work-explore	fork
work-brainstorm	manual only — run /work-brainstorm	inline
work-pr	manual only — run /work-pr	inline
dev-api	"API", "endpoint", "REST"	inline

Additional skills

Skill	Trigger	Context
dev-flutter	"Flutter", "widget", "BLoC"	inline
dev-supabase	"Supabase", "auth", "RLS"	inline
dev-react-perf	"React perf", "re-render", "memo"	inline
ops-docker	manual only — run /ops-docker	inline
ops-ci	manual only — run /ops-ci	inline
ops-database	manual only — run /ops-database	inline
ops-monitoring	manual only — run /ops-monitoring	inline
doc-generate	manual only — run /doc-generate	fork
doc-changelog	manual only — run /doc-changelog	fork
dev-refactor	"refactor", "clean code", "restructure"	inline
dev-error-handling	"error handling", "exceptions", "error boundary"	inline
dev-graphql	"GraphQL", "resolver", "schema"	inline
ops-mobile-release	"App Store", "Play Store", "Fastlane"	inline
data-pipeline	"ETL", "Airflow", "dbt"	inline
qa-perf	"optimize", "latency", "TTFB"	inline
qa-e2e	"E2E", "Playwright", "Cypress", "user journey"	inline
feature-flags	"feature flag", "A/B test", "progressive deployment"	inline
ops-infra-code	"Terraform", "IaC", "OpenTofu", "module", "state"	inline
ops-proxmox	"Proxmox", "PVE", "Proxmox VM", "LXC", "PBS"	inline
ops-opnsense	"OPNsense", "firewall", "NAT", "DHCP", "Unbound"	inline
qa-tech-debt	"technical debt", "tech debt", "refactoring priority"	fork
ops-standup	"standup", "briefing", "what happened"	fork
ops-ci-fix	"ci broken", "fix ci", "workflows failing"	inline
qa-design	"design audit", "UI/UX", "user interface"	fork
api-mocking	"mock API", "MSW", "test without backend"	inline
state-management	"state", "Redux", "Zustand", "store"	inline
dev-document	"PDF export", "invoice", "report job", "generate DOCX/XLSX/PPTX" from app code	inline
growth-cro	"conversion", "CRO", "signup flow", "onboarding", "paywall"	inline
parallel-agents	"parallel", "concurrent", "fan-out", "multi-agents"	inline
agent-teams	"agent team", "swarm", "agent team", "parallel agents"	inline
session-handoff	"handoff", "resume", "session transfer", "context"	inline
git-worktrees	"worktree", "parallel dev", "simultaneous branches"	inline
qa-chrome	manual only — run /qa-chrome	inline
dev-frontend-design	"UI design", "landing page", "art direction", "fonts"	inline
dev-shadcn	"shadcn", "shadcn/ui", "Radix", "React components"	inline
dev-nextjs	"Next.js", "App Router", "Server Components", "RSC", "Server Actions"	inline
dev-auth	"auth", "login", "signup", "OAuth", "better-auth", "NextAuth", "Lucia", "2FA"	inline
dev-prisma	"Prisma", "schema.prisma", "migrate", "ORM" — only while no prisma-* vendor skill is installed	inline
dev-i18n	"i18n", "l10n", "translation", "locale", "next-intl", "react-i18next", "vue-i18n", "flutter_localizations"	inline
writing-skills	"create skill", "new skill", "write a skill"	inline
web-scraping	"scrape", "crawl", "extract web", "Firecrawl", "structured data"	fork
work-quick	"quick", "fast", "rapid" — trivial change (< 50 LOC, 1-3 files)	inline
work-batch	"batch", "backlog", "PRD", "user stories in series" — sequential execution	fork

Skills Configuration

Each skill's frontmatter may carry:

• allowed-tools: Tools pre-approved for the skill's turn (grants, never restricts). The foundation declares none — see Pre-approving tools
• context: absent (the default) = the skill runs inline, in the conversation; context: fork = a sub-agent that sees none of it, for a self-contained job only (with background: false)

Skills are triggered automatically by Claude based on context — except those marked manual only in the tables above, which carry disable-model-invocation: true and can be started by you alone.

Skill overrides (CLI 2.1.129+)

Bundled skills can be selectively suppressed at the settings level via the skillOverrides key in .claude/settings.json or .claude/settings.local.json. Three modes are accepted:

Mode	Effect
off	The skill does not load and is not invocable.
user-invocable-only	Automatic triggering is disabled; the skill remains available via explicit /skill-name invocation.
name-only	The skill name is preserved in the catalog without loading its body — useful when the name is referenced elsewhere but a vendor alternative should take over the work.

When to reach for it: a vendor-published skill duplicates the depth of a bundled one (e.g., a vendor framework's official skill replaces the foundation's general-purpose equivalent). Setting the bundled skill to user-invocable-only lets the vendor skill auto-trigger while keeping the bundled one as a manual fallback.

Note on the foundation's preset filter — each preset's foundation.skills.drop[] (blacklist) or foundation.skills.keep[] (whitelist, mutually exclusive with drop[], enforced by validate-presets.sh) array operates at a different layer: it filters skills out at claude-base init time, before they are ever installed in the project. skillOverrides operates at session start on already-installed skills. The two mechanisms are complementary and can be combined.

Refer to the upstream Claude Code changelog for the canonical JSON shape and any future modes added to the setting; this section names the modes shipped at 2.1.129 and describes their intent, not their evolving syntax.

Skills Best Practices

Size and Budget

• SKILL.md < 500 lines (offload bulky content into reference files)
• Description budget: 15k characters max (SLASH_COMMAND_TOOL_CHAR_BUDGET)
• Use support files: examples/, scripts/, reference.md

Complete frontmatter

---
name: my-skill
description: Short description of the skill
context: fork                     # only for a self-contained job; omit = inline
background: false                 # with fork: wait for the result in the same turn
disable-model-invocation: true   # Do not trigger automatically
user-invocable: false             # Background-only skill
argument-hint: "[description]"    # Hint for arguments
model: sonnet                     # forked skills only: inline, it switches the session model
agent: my-agent                   # Associated agent
hooks:
  PreToolUse:
    - matcher: "Bash"
      hooks:
        - type: command
          command: "echo validation"
---

Variable substitutions

Variable	Description
$ARGUMENTS	All arguments passed to the skill
$ARGUMENTS[N]	Argument at index N (0-based)
$N	Shortcut for $ARGUMENTS[N]
${CLAUDE_SESSION_ID}	ID of the current session

Dynamic Context Injection

Inject dynamic content into a skill with the syntax:

!`command`

Example: !`cat package.json | jq .scripts` injects the npm scripts into the skill's context.