Claude Code skills are folders of instructions, scripts, and reference files that Anthropic's coding agent loads on demand to handle specialized tasks. On GitHub, they live in a growing ecosystem of official repositories, community marketplaces, and single-purpose plugins, all built around a shared SKILL.md format.
What a Claude Code skill actually is
A skill is a folder with a SKILL.md file at its root. That file contains YAML frontmatter with two required fields and a body of instructions the model follows when the skill is active.
---
name: my-skill-name
description: A clear description of what this skill does and when to use it
---
# My Skill Name
Instructions Claude will follow when this skill is active.
## Examples
- Example usage 1
- Example usage 2
## Guidelines
- Guideline 1
- Guideline 2
Minimal SKILL.md
Alongside SKILL.md, a skill can ship optional scripts/ (executable helpers, often Python or shell), references/ (longer documentation loaded only when needed), and assets/ (templates, fixtures, or images). Claude loads the frontmatter up front and pulls in the rest on demand, which keeps token usage low until a task actually needs the detail.
The name must be lowercase with hyphens, and the description should explain both what the skill does and when to use it. That description is what Claude reads to decide whether to activate the skill, so vague wording leads to skills that never fire.
The official Anthropic skills repository
The canonical starting point is github.com/anthropics/skills. It hosts Anthropic's demonstration skills plus the production document-handling skills that power Claude's file creation features.
The repo is organized into three top-level areas:
| Path | Contents |
|---|---|
./skills | Example skills for creative work, development, enterprise communication, and document handling |
./spec | The Agent Skills specification |
./template | A starter skill you can copy |
The document skills — skills/docx, skills/pdf, skills/pptx, and skills/xlsx — are source-available rather than open source, and they're the same ones running behind Claude's file creation feature. Most other skills are Apache 2.0 licensed.
For background on the system itself, Anthropic's engineering post Equipping agents for the real world with Agent Skills explains the design intent, and the Claude support center covers usage details for end users.
Installing skills in Claude Code
Claude Code treats GitHub repositories as plugin marketplaces. Adding one registers the repo so you can browse and install the skill bundles it defines.
Step 1: Open Claude Code and register the marketplace. For the official repo, run /plugin marketplace add anthropics/skills. Any public GitHub repo with a valid marketplace manifest works the same way.
Step 2: Install a plugin bundle. You can either run /plugin install document-skills@anthropic-agent-skills directly, or type /plugin, choose Browse and install plugins, pick the marketplace, select a plugin, and confirm.
Step 3: Invoke the skill by mentioning it in a prompt. After installing document-skills, a message like "Use the PDF skill to extract form fields from path/to/file.pdf" will trigger the right skill without additional configuration.
For Claude.ai (paid plans) and the API, the same skills are available through the product UI and the Skills API quickstart, respectively. The SKILL.md format is identical across all three surfaces.
Community marketplaces worth knowing
Beyond the official repo, several large community collections have emerged. They generally follow the same plugin-marketplace pattern, so installation feels familiar once you've done it once.
| Repository | Focus | Install command |
|---|---|---|
anthropics/skills | Official examples and document skills | /plugin marketplace add anthropics/skills |
alirezarezvani/claude-skills | 200+ skills across engineering, product, marketing, compliance, C-suite advisory | /plugin marketplace add alirezarezvani/claude-skills |
daymade/claude-code-skills | ~48 workflow skills including skill-creator, github-ops, and document conversion | /plugin marketplace add daymade/claude-code-skills |
rohitg00/awesome-claude-code-toolkit | Curated agents, skills, hooks, commands, and MCP configs | /plugin marketplace add rohitg00/awesome-claude-code-toolkit |
levnikolaevich/claude-code-skills | Seven installable plugins covering Agile workflow, audits, and bootstrapping | /plugin marketplace add levnikolaevich/claude-code-skills |
ComposioHQ/awesome-claude-skills | Awesome-list style catalog linking to skills across the ecosystem | Browse on GitHub; install skills individually |
The GitHub topic page at github.com/topics/claude-code-skills aggregates repositories tagged with claude-code-skills, which is useful for discovery when you want something narrower than a mega-marketplace.
Skills, subagents, and plugins — how they differ
The terminology in this space overlaps, which causes confusion. A rough mental model:
| Concept | What it is | Scope |
|---|---|---|
| Skill | Instructions plus optional scripts and references for executing a specific task | Single domain, loaded on demand |
| Subagent | A delegated worker Claude spawns to handle part of a task in isolation | Task-scoped, often used for parallel work |
| Plugin | A packaged bundle distributed via a marketplace, typically containing one or more skills | Distribution unit |
| Persona | A curated combination of skills and style guidance that defines how an agent thinks and communicates | Cross-domain identity |
Skills are the most portable piece because the SKILL.md specification is the same standard referenced at agentskills.io. Several community toolchains convert the same skill folders to formats used by OpenAI Codex, Gemini CLI, Cursor, Aider, Windsurf, and other coding agents, so a well-written skill often runs in more than one environment without changes.
Creating your own skill
The Anthropic template folder is the simplest starting point, but most people get further faster with a meta-skill like skill-creator, which ships in both the official repo and community forks.
Step 1: Create a folder named after your skill using lowercase letters and hyphens. Inside it, add a SKILL.md with the two required frontmatter fields. Keep the description specific about both the task and the trigger conditions.
Step 2: Write instructions in the body that read like a short runbook. Cover the decision points, the exact commands or API calls, and the expected outputs. If a procedure is long, move the detail into references/ and link to it so Claude only loads it when needed.
Step 3: Add scripts to scripts/ only when deterministic code beats instructing the model. Standard-library Python and shell scripts are common because they avoid dependency management. Document how to call each script in SKILL.md.
Step 4: Test the skill locally by copying the folder into ~/.claude/skills/ and asking Claude Code to perform the task it's designed for. If the skill doesn't activate, the description is usually the problem — make it more explicit about the triggering user intent.
Step 5: Package for distribution by committing the folder to a public repo, adding a .claude-plugin/marketplace.json manifest, and sharing the install command. Anthropic's plugin marketplaces documentation covers the manifest format.
Design patterns that hold up in production
A few patterns repeat across the best-regarded community skills:
Progressive disclosure. The SKILL.md is short and decisive. Long procedures, API references, and edge cases live in references/ files that Claude reads only when the task requires them. This keeps context windows lean for simple requests.
Deterministic scripts for fragile work. Tasks like parsing complex spreadsheets, scanning for secrets, or generating valid migrations are handled by bundled scripts rather than free-form model output. The skill instructs Claude to call the script and interpret its result.
Explicit triggers in the description. Descriptions that name the specific user phrases or file types that should activate the skill fire more reliably than generic ones. "Use when the user asks to convert a .docx to markdown" works better than "Document conversion utilities."
Zero-dependency tooling. Many production skills deliberately limit scripts to the Python standard library or POSIX shell so they run anywhere without pip install steps. This matters more than it sounds — dependency failures are one of the top reasons skills break for other users.
Security review before installation. Community projects increasingly ship security auditors that scan a skill folder for command injection, data exfiltration, and prompt injection risks before you install it. Running one of these against an unfamiliar skill takes seconds and catches real problems.
Confirming a skill is working
After installation, Claude Code shows installed plugins through the /plugin command. If a skill doesn't appear to activate when expected, the common causes are:
- The plugin is installed but not enabled in
settings.jsonunderenabledPlugins. - The SKILL.md description is too vague for Claude to match the user's request.
- The marketplace cache is stale — removing and re-adding the marketplace refreshes it.
- The skill folder has invalid YAML frontmatter, causing Claude Code to skip loading it with a warning.
You can verify a skill loaded correctly by asking Claude directly, for example: "Which skills do you currently have available?" It will list the active skills along with their descriptions, which also makes it easy to check whether your own SKILL.md is being read the way you intended.
Where the ecosystem is heading
The SKILL.md format has effectively become a portable standard. The same folder structure runs in Claude Code, Claude.ai, the Claude API, and — through conversion scripts — a growing list of third-party coding agents. That portability is why the GitHub ecosystem has scaled so quickly: a skill written once can reach users across multiple tools without rewrites.
For most people, the practical path is to start with Anthropic's official repo to understand the conventions, install one community marketplace that matches your work, and write your first skill only when you hit a repeatable task that the existing catalog doesn't cover. The barrier to entry is low — a folder, a markdown file, and a clear description are enough to extend Claude Code in ways that stick.