Agents

Agent Skills

Reusable packages of instructions, scripts and reference files that an agent loads only when a task calls for them.

ONE SKILL = ONE FOLDERName + description01always in context · a few dozen tokensSKILL.md instructions02loaded when the task is relevantScripts, templates, files03opened only when neededProgressive disclosure: an agent can know dozens of skills and fill its context only with the one it uses.

swipe to see the whole diagram →

MEmehmeterkek.com/glossary/agent-skills

In plain terms

A new hire is capable but does not know how your company does things. You hand them a binder: how we write proposals, how we close the books. A skill is that binder for an agent: a folder holding the procedure and the files it needs, opened when the matching task comes up.

Why it matters

Skills turn know-how that lives in people's heads into something an agent applies the same way every time, and they can be versioned, reviewed and shared like any other document. They also keep costs down, since instructions the agent is not using stay out of its context.

Example

A “quarterly report” skill holds the house template, the rules for rounding and currency, and a script that builds the charts. An analyst asks the agent for the Q3 report; the agent recognises the match, loads the skill and produces the report in the company format.

Most often confused with

Agent Skills vs. MCP (Model Context Protocol)

Agent SkillsTeaches the agent how to do something
MCP (Model Context Protocol)Gives the agent access to something

MCP connects the agent to systems: it supplies tools and data. A skill supplies procedure: the steps, standards and judgment for a task. They complement each other, and a skill often explains how to use an MCP server's tools well.

Origin: Introduced by Anthropic in October 2025 and later published as an open specification.

Under the hood

A skill is a folder with a SKILL.md file: YAML front matter (name and description) followed by Markdown instructions, optionally alongside scripts and resource files. Loading is progressive: only names and descriptions sit in the context by default; the body is read when the model judges the skill relevant; bundled files are opened or executed on demand. The description therefore decides whether a skill triggers. Skills need a code-execution environment to run their scripts and should be reviewed like code, since they can carry instructions and executables from third parties.

Written by Mehmet Erkek · Last updated: