What Are Claude Skills?

Claude Skills are reusable instruction packages that teach Claude or Claude Code how to perform a specific workflow. Most skills are built around a SKILL.md file with metadata and Markdown instructions.

Instead of explaining the same process every time — how to review code, write docs, create a campaign brief, or inspect a security risk — you install a skill once and let Claude load it when the task is relevant.

The short version

A skill usually contains:

my-skill/
├── SKILL.md
├── references/
├── scripts/
└── examples/

The required piece is SKILL.md. It usually includes YAML frontmatter:

---
name: code-reviewer
description: Reviews code for bugs, security issues, and maintainability risks.
---

…and body instructions:

# Code Review Process

1. Understand the change.
2. Identify correctness risks.
3. Check security-sensitive surfaces.
4. Suggest minimal fixes.
5. Summarize verification steps.

Why skills matter

Skills turn agent behavior into something closer to a repeatable system:

Consistency: the same checklist runs every time.
Portability: a workflow can move between projects.
Focus: the agent loads task-specific context instead of global instructions.
Team reuse: one good workflow can become a shared team standard.

Skills vs prompts

A prompt is usually one-off. A skill is packaged for reuse.

Prompt	Skill
Pasted manually	Installed once
Easy to lose	Reusable across sessions
Often broad	Usually task-specific
No file structure	Can include references, examples, and scripts

Skills vs MCP servers

Skills and MCP servers solve different problems.

Skills teach the agent how to do a workflow.
MCP servers give the agent tools or data it can call.

A strong setup often uses both. For example: an MCP server exposes GitHub issues, while a skill teaches Claude how to triage, label, and summarize them.

Are Claude Skills safe?

A skill is powerful because it can shape the agent's behavior. That also means you should inspect it before installing.

Review:

the source repository
the author
install commands
shell scripts
network access
credential handling
hidden or obfuscated files
whether the skill asks for more permissions than it needs

Skills Directory scans skills and surfaces security grades to make that review easier, but no automated scan replaces human judgment.

Good skills are narrow

The best skills do one job well:

code review
systematic debugging
writing release notes
creating a landing page brief
checking accessibility
planning a database migration

Broad skills that try to control all agent behavior often conflict with project instructions and other skills.