Agent Skills: What SKILL.md Is, Where It Works, and Where to Get Them (2026)

Agent Skills are folders of instructions that an AI agent loads only when a task needs them. Each one has a SKILL.md file: a short header with a name and a description, then plain Markdown instructions, plus optional scripts and templates. Anthropic created the format for Claude and published it as an open standard, and it now works in Claude, Claude Code, OpenAI Codex and GitHub Copilot in VS Code, among others. Write a skill once and the same folder works in every agent that supports the format.

Checked against the official docs on 21 October 2026 · By the KissMySkills team

What is an agent skill?

An agent skill is a set of instructions packaged so an AI agent can find it and use it by itself. You don't paste it into every chat. You install it once, and the agent decides when it is relevant. The Agent Skills specification defines it as a directory containing, at minimum, a SKILL.md file:

That's all there is. No SDK, no account, no build step. A skill is plain files you can open in a text editor, put in version control and share as a ZIP.

The SKILL.md format

SKILL.md starts with YAML frontmatter between two --- lines, followed by Markdown instructions. The spec requires two fields:

A minimal, valid skill looks like this:

The description does most of the work. "Helps with reports" will never trigger. A description that names the task and the words people actually type will.

How do agents load skills?

Agents load skills in three steps, which the spec calls progressive disclosure:

1. Discovery. At startup the agent reads only the name and description of every installed skill.
2. Activation. When your request matches a description, the agent reads the full SKILL.md.
3. Execution. It follows the instructions and opens bundled files or runs scripts only if it needs them.

This is why you can install dozens of skills without filling the context window. Until a skill is used, it costs a line or two of text.

A finished agent skill

Serge - SEO Specialist AI Skill

$19this skill, one-time vs $1,500an SEO agency, per month

A standard SKILL.md package: keyword clusters, technical audits and content briefs, with the rules for what to ask first and what never to guess. Upload it to Claude, drop it into a skills folder, or load it through the connector.

See the SEO skill →

Where do agent skills work?

The same folder works in every tool below. What changes is where you put it.

Folder names and menus change often. If a step above doesn't match what you see, the vendor's own docs (linked in the table) are the source of truth.

Agent skills vs prompts vs MCP

• A prompt is a one-off message. You retype or paste it every time.
• A skill is know-how the agent keeps on hand and loads by itself: how to do a job, in what order, in what format.
• MCP (Model Context Protocol) gives the agent access: tools and data in other apps, such as your CRM, Notion or Google Sheets. In Claude these are called connectors.

They work together. A skill tells the agent how to write a sales follow-up; an MCP connector lets it read the deal from your CRM. More in Claude Skills vs MCP and AI agents vs skills vs prompts.

Where to get agent skills: the marketplaces

There are three kinds of source, and they solve different problems:

• First-party skills. Anthropic publishes a small, reviewed set (the anthropics/skills repository and built-in skills for Word, Excel, PowerPoint and PDF). OpenAI publishes skills for Codex. Free and reliable, but few of them, and mostly about documents and code.
• Free indexes. Sites such as SkillsMP, skills.sh and claudeskills.info collect SKILL.md files from public GitHub repositories. They are huge and free. Quality varies from excellent to abandoned, and you review each skill yourself.
• Written catalogs. Marketplaces that write and maintain their own skills, like KissMySkills, built around jobs rather than code: an SEO specialist, a CFO, a recruiter, a copywriter. You pay, and in return someone is responsible for the file.

We compare the main options, with prices checked on each vendor's site, in Best Claude skills marketplaces.

Are agent skills safe to install?

A skill that is only Markdown instructions can't do anything your agent couldn't already do. A skill with a scripts/ folder can run code on your machine when the agent uses it. Before installing a skill from an index or a stranger's repository:

1. Read SKILL.md from start to finish. Look for instructions to send data anywhere, or to ignore your instructions.
2. Open every file in scripts/. If you can't tell what a script does, don't install it.
3. Check the repository: who maintains it, when it last changed, and what license it uses. GitHub stars tell you it's popular, not that it's safe.
4. In coding agents, keep tool approvals on for skills you haven't checked.

How to write your own agent skill

1. Pick one job. "Weekly status report" works. "Help with work" doesn't.
2. Write the description first. Name the task and the words a user would type. It decides when the skill runs.
3. Write the steps you'd give a new colleague: what to ask first, the order of work, the output format, and what never to guess.
4. Move long reference material into references/ and link it from SKILL.md, so it loads only when needed.
5. Test it on three real tasks. If you had to correct the same thing twice, add a rule.

To skip the blank page, the free Claude Skill Builder turns a plain-English description of a job into a SKILL.md you can download.

Related guides

• What Are Claude Skills? The Complete Guide
• How to install a Claude skill
• Best Claude Code skills in 2026
• What Is MCP (Model Context Protocol)?

KissMySkills is an independent marketplace and is not affiliated with Anthropic, OpenAI, GitHub or Microsoft. Claude, Codex and Copilot are trademarks of their owners.