Field guide · Agent Skills
Skills went from an Anthropic feature to an open standard with its own package manager in about a year. Here is the map: the directories, the big repositories, the plugin marketplaces, and the handful of articles actually worth your time.
A folder. Inside it, a required SKILL.md — YAML frontmatter plus markdown instructions — and optionally scripts/ for executable code, references/ for docs loaded on demand, and assets/ for templates.
my-skill/
├── SKILL.md # required: frontmatter + instructions
├── references/ # loaded only when needed
├── scripts/
└── assets/The mechanism that makes this cheap is progressive disclosure: the agent first sees only each skill's name and description — roughly a hundred tokens — and pulls the full body in only when the description matches what you asked for. That is why the description field does more work than the rest of the file combined.
Anthropic introduced the format in October 2025 and released it as an open standard in December 2025. It is now read by Claude Code, claude.ai, the Claude API, OpenAI Codex, Cursor, Gemini CLI, Antigravity, Windsurf, and others — which is why a "Claude skill" is increasingly just a skill.
Start here rather than in GitHub search. These index across repos and give you an install command.
Vercel · the closest thing to npm for skills
1.3M+ total installs reported across the ecosystem
A searchable catalog of published skills with install counts, backed by the vercel-labs/skills CLI. Browse by category, author, or popularity; install into any of ~17 supported agents. The fastest route from "I need a thing that does X" to a working skill.
npx skills add <owner/repo>Community · plugin-marketplace index
Indexes Claude Code plugin marketplaces rather than bare skills — useful when you want a bundle (skills + agents + hooks + MCP) instead of a single file, and when you want to see what a marketplace ships before adding it.
Community · browsable directory
A web front end over the awesome-list ecosystem. Lower ceremony than cloning a repo to read a README, and it surfaces the same entries with a bit more description.
The open standard itself
The spec skills conform to. Worth ten minutes if you plan to write and publish, because it tells you which frontmatter fields are portable across agents and which are Claude Code extensions.
On the install numbers. skills.sh reports figures in the millions for its top entries. Treat those as directional. An "install" there is a CLI fetch, not a human deciding a skill was good — a skill high on that list is popular and discoverable, which is not the same as correct for your repo.
Two repos, both first-party, and the place to start if you'd rather not audit a stranger's shell scripts.
Anthropic · reference skills + the spec
≈174k stars
Anthropic's own skills, organized into Creative & Design, Development & Technical, Enterprise & Communication, and Document Skills — the docx / pdf / pptx / xlsx set is the one most people arrive for. Also carries ./spec and a ./template to start from.
/plugin marketplace add anthropics/skills
/plugin install document-skills@anthropic-agent-skills
/plugin install example-skills@anthropic-agent-skillsAnthropic · curated plugin directory
≈36k stars
The managed directory of Claude Code plugins — /plugins for Anthropic's own, /external_plugins for vetted third parties. Includes skill-creator, a skill that writes skills: describe a procedure or paste an SOP and it generates the folder, the routing description, and the body, then A/B tests the result and reports token overhead.
/plugin install skill-creator@claude-plugins-official
# or browse: /plugin > Discover| Repository | Stars | What it is |
|---|---|---|
| ComposioHQ/ awesome-claude-skills |
≈74.5k | The biggest list by a distance — 1000+ skills across 11 categories, from document processing to security to assistive tech, plus 78 SaaS integrations via Composio. Cross-agent, not Claude-only. |
| travisvn/ awesome-claude-skills |
≈15k | The original community list, and still the better-edited one. Official and community skills separated, plus creation tools, learning resources, a security section, and a FAQ. Read this one first. |
| BehiSecc/ awesome-claude-skills |
— | A third curated list. Smaller, some entries the larger lists miss. |
| shanraisshan/ claude-code-best-practice |
— | Not a skills list — a running collection of practice notes and tips from Anthropic engineers. Useful alongside the directories. |
Star counts on curated lists measure the list, not the skills. A 74k-star awesome-list is 74k people bookmarking a README. The individual skill inside it may have been written once, in an afternoon, and never run against your stack.
npx skills add <owner/repo> # interactive picker
npx skills add vercel-labs/skills # e.g. find-skillsNot installed globally — npx always pulls the current version. It will ask which agent to install into and drop files in the right place for it.
/plugin marketplace add <owner/repo>
/plugin install <plugin>@<marketplace>The right choice when you want skills, subagents, hooks, and MCP servers arriving together and updating together. Plugin skills are namespaced as /plugin-name:skill-name, so they never collide with yours.
.claude/skills/<name>/SKILL.md # commit it — the team gets it on clone
~/.claude/skills/<name>/SKILL.md # personal, every projectUnglamorous and usually correct. A skill you have read line by line and checked into your own repo has no supply chain.
CLAUDE_CODE_SYNC_SKILLS=1 claude -p "List your skills"
# lands in ~/.claude/skills/synced/Synced skills deliberately do not execute inline !`shell` blocks. That restriction is a feature.
Frontmatter reference, progressive disclosure, argument substitution, dynamic context injection. Everything else on this list is commentary on this page.
Thariq Shihipar, Anthropic technical staff, on how Anthropic uses "hundreds" of skills internally and the nine categories they fall into. The single most useful thing written about skill design rather than skill syntax. Coverage linked; the original is on the Claude blog.
What is portable across agents and what is Claude Code-specific. Read before publishing anything.
A clean walkthrough of the authoring guidance, if you want it in one page rather than across the docs.
Opinionated shortlist with reasoning, rather than a dump of every repo that exists.
Another shortlist. Useful mainly for the overlap — a skill both lists name is worth a look.
Short survey of the big repos, useful as a sanity check on this page.
The registry side of the story from the people who built the CLI.
A skill is instructions plus, frequently, executable code, running with your agent's permissions in your repository. The install is one command; the blast radius is not.
| Check | Why |
|---|---|
Read SKILL.md end to end | It is usually under 500 lines. If it is longer than that, that is itself a finding |
Open scripts/ | This is where arbitrary code lives. A skill with no scripts is a much smaller decision |
Look for !`…` blocks | Inline shell runs before Claude sees the skill, every invocation, with no permission prompt |
Check allowed-tools | It pre-approves tools without prompting. A broad Bash(*) here is a broad grant |
Check disable-model-invocation | Absent means Claude can fire it on its own, whenever the description seems to match |
Read the description critically | An over-broad description means the skill loads constantly and eats context you wanted for code |
Run /context after installing | Shows what the skill listing now costs you every single session |
Run /skill-doctor monthly | Finds the ones you installed, never used, and are still paying for |
The cost of a skill is not the install. It is the context it occupies in every session, forever, whether it fires or not.
Three settings are worth knowing when a collection gets out of hand: skillOverrides hides or collapses one skill without editing its file, disableBundledSkills turns off everything shipped with Claude Code, and disableSkillShellExecution kills inline shell across all of them at once.
