What Is An AI Skill?

Think of a skill as reusable operating instructions for an agent, not as a server process. A good skill narrows one task, defines expectations, and sometimes ships helper scripts or templates. It is closer to a playbook than to an API.

That is why skills are useful even when they contain no executable server. They can still improve planning, output structure, review discipline, and tool usage patterns dramatically.

• A primary instructions file such as SKILL.md.
• Optional scripts, templates, examples, or reference files.
• Optional metadata that tells a host application how to display or invoke the skill.
• Optional dependency declarations, especially in platforms that support richer skill packaging.

That structure matters for review. A skill that looks harmless at the prompt layer can still ship scripts or helper files that widen its effective permissions.

Use a skill when you want repeatable guidance, not a long-running tool transport. Examples include code review standards, threat-model checklists, release playbooks, data extraction workflows, or organization-specific documentation conventions.

If the capability needs live access to a database, browser, design tool, or SaaS API, that usually points to an MCP server instead.