What Is a Claude Skill?

The short version

A Claude Skill packages a repeatable way of doing work. Instead of pasting the same checklist, style guide, file-handling rules, or debugging procedure into every conversation, you place that material in a skill. Claude can then discover and load the skill when the current request matches its scope.

Anthropic’s current definition is deliberately broader than a prompt file: a skill can include instructions, executable scripts, reference material, templates, and other resources. The detailed instructions are loaded only when needed, which keeps unrelated material out of the active context.

What is inside a skill?

The basic unit is a directory containing SKILL.md. Its YAML frontmatter identifies the skill and describes when it should activate. The Markdown body contains the actual workflow.

A minimal structure looks like this:

my-skill/
├── SKILL.md
├── references/       optional documentation
├── scripts/          optional deterministic helpers
└── assets/           optional templates or media

The description matters because it is part of skill discovery. A vague description such as “helps with documents” gives Claude little information about when the skill belongs in a task. A useful description names the expected input, output, and trigger conditions.

How Claude loads a skill

Anthropic describes the loading model as progressive disclosure:

1. Claude can inspect lightweight skill metadata without loading every instruction.
2. When a request matches the skill, Claude loads SKILL.md.
3. Detailed references, scripts, or assets can be opened only when the workflow requires them.

This makes a skill suitable for procedures that are valuable repeatedly but too detailed to keep in permanent project instructions.

Where Claude Skills run

Skills currently appear across several Anthropic surfaces, but installation and management differ:

Surface	How skills are used
Claude.ai and Desktop	Browse supported packages or upload custom skills in Customize > Skills. Code execution and file creation must be enabled for workflows that need them.
Claude Code	Store skills under .claude/skills/, a user skill directory, or install a plugin that contains skills. Claude can load them automatically or expose them as slash-invoked workflows.
Claude API	Use prebuilt or uploaded skills with the code-execution tool.
Team and Enterprise	Owners can provision approved skills across the organization.

Availability and controls can vary by plan and organization policy. Check Anthropic’s current Skills overview and usage guide before planning a deployment.

Anthropic now also provides a unified directory for browsing skills, connectors, and plugins from the Customize area. That product directory is the supported installation surface; independent directories such as Agent Skill Picks add cross-repository source and maintenance checks. See the current marketplace and directory guide for the distinction.

For setup instructions, continue with How to Use Claude Skills. For project folders and plugin commands, use the Claude Code Skills guide.

Claude Skill examples

Good skill candidates have a clear trigger and a repeatable result:

• Produce a Word document with an established structure and verification process.
• Build an Excel workbook while preserving formulas and formatting conventions.
• Review a pull request using a team’s evidence and severity rules.
• Apply a brand system to presentations and marketing assets.
• Debug failures with a fixed reproduce, isolate, verify, and regress workflow.
• Create another skill using the required frontmatter and folder structure.

Browse the rule-curated Claude Skills directory for source-backed examples, or start with the official Anthropic creator page.

When to use a skill

Use a skill when all or most of these are true:

• The workflow repeats across projects or conversations.
• Order matters and skipping a step causes predictable errors.
• Reference files, templates, or scripts improve the result.
• The task has recognizable trigger conditions.
• The instructions are useful only for this class of task, not every conversation.

For example, “make this paragraph clearer” probably needs only a prompt. “Convert every customer interview into the same evidence table, apply our research taxonomy, flag unsupported conclusions, and export a standard brief” is a much stronger skill candidate.

When not to create a skill

Do not make a skill merely to store one clever prompt. A normal prompt or project instruction is usually better when the task is one-off, the procedure changes every time, or the rule must remain active in every conversation.

Avoid installing a third-party skill until you have inspected its instructions, scripts, dependencies, and network behavior. Anthropic warns that skills can contain code or instructions that install software, creating prompt-injection and data-exfiltration risks. Start with trusted sources and test on non-sensitive inputs.

Skills compared with related features

Feature	Primary job	Use it when
Skill	Reusable knowledge and procedure	Claude needs a repeatable way to complete a task
Plugin	Package and distribute extensions	A team wants to install skills, agents, hooks, or MCP configuration together
MCP	Connect external tools and data	Claude needs to query or act in another system
Agent	Run a delegated reasoning loop	The work benefits from a separate specialist context
Hook	React automatically to lifecycle events	A deterministic check or action must run at a specific event
Project instructions	Persistent project context	A rule should apply throughout work in one project

These features are composable. An installed plugin can provide a skill; that skill can teach Claude how to use tools from an MCP server; and the workflow can delegate a focused step to an agent.

Read the focused comparisons for Skills vs Plugins, Skills vs MCP, Skills vs Agents, and Skills vs Commands.

How to judge a skill before using it

Check more than the repository’s star count:

1. Confirm the exact folder contains a SKILL.md, not merely a link in an aggregator.
2. Read the description and verify that the trigger matches your task.
3. Inspect scripts, dependencies, permissions, and network instructions.
4. Check the license and the date the source was last maintained.
5. Run a representative, non-sensitive input and inspect the output.

Agent Skill Picks verifies public source signals and clearly separates those profiles from hands-on reviews. The methodology explains the quality and indexing gates.

Primary sources

• Anthropic: What are skills?
• Anthropic: Use skills in Claude
• Anthropic: Browse skills, connectors, and plugins
• Claude Code: Extend Claude with skills
• Agent Skills open standard