Black Friday sale: 30% off Flywrite

Agent Skills are an open format for reusable instructions you give an AI agent: each skill is a folder with a SKILL.md file that the agent reads when a task calls for it, and Claude Code, Codex, Cursor and other agents all read the same format.

What it looks like

A skill is a folder with a SKILL.md file in it. It can hold more files beside it, such as a references/ folder the agent opens only when the steps call for it. Here's a writer's skill that checks a draft against a house style sheet:

house-style/
├── SKILL.md
└── references/
    └── style-sheet.md

SKILL.md is Markdown with front matter on top. Two fields are required: the name, which matches the folder name and uses only lowercase letters, numbers and hyphens, and the description, which says what the skill does and when to use it:

---
name: house-style
description: Check a draft's house style before it goes out.
---

# House style check

1. Read references/style-sheet.md.
2. Check headings, numbers and dates against it.
3. Flag any word the style sheet bans.
4. List each change to make, with its line.

The agent reads a skill in three stages, so a long style sheet stays out of the way until a task needs it:

  1. When a session starts, it loads only each skill's name and description.
  2. When you ask for a style check, it reads the whole SKILL.md.
  3. It opens style-sheet.md only when the skill's first step tells it to.

One skill, every agent

Make a skill when you keep pasting the same checklist or procedure into every session: a style check, a pre-send pass, the steps that turn notes into an outline. The format is the same in every agent that supports it, but each agent looks in its own folders:

Agent Project skills Personal skills Run one with
Claude Code .claude/skills/ ~/.claude/skills/ /name
Codex .agents/skills/ ~/.agents/skills/ $name, or pick from /skills
Cursor and Cursor Agent .agents/skills/ or .cursor/skills/, plus .claude/skills/ and .codex/skills/ The same four folders in ~/ The / menu
GitHub Copilot .github/skills/, .claude/skills/ or .agents/skills/ ~/.copilot/skills/ or ~/.agents/skills/ Loads it when relevant

Each of them can also pick a skill on its own when your request matches its description. Which of Copilot's folders work depends on where you use it, such as Copilot CLI or VS Code.

To use one skill in all four, keep it in .agents/skills/, where Codex, Cursor and Copilot read it, and link it into .claude/skills/ for Claude Code, which follows a linked skill folder. From the project folder:

mkdir -p .claude/skills
ln -s ../../.agents/skills/house-style .claude/skills/house-style

If you don't use Codex, .claude/skills/ alone covers the other three, since Cursor and Copilot read it too.

What trips people up:

  • The description decides when the agent picks the skill. Write it in the words you'd use to ask for the task.
  • Extra fields don't travel. For a skill you share, stick to the standard's fields: name, description, license, compatibility, metadata and allowed-tools. Extras such as Claude Code's disable-model-invocation work only in their own agent.
  • A skill can run code. The format allows a scripts/ folder of code for the agent to run, so add skills only from people you trust.

Agent Skills, AGENTS.md and MCP

Three things shape what your agent knows and can do, and each has its own job:

What it gives the agent When it loads A writer's example
AGENTS.md, or CLAUDE.md in Claude Code Standing instructions Before every session's work Your voice, the folders to leave alone
A skill A procedure When a task calls for it The style check before an issue goes out
An MCP server A connection to another app When a task needs that app Your Notion notes or your calendar

They work together. A skill can spell out the steps for a job that uses a connected app, and Anthropic describes skills as complementing MCP servers.

Anthropic introduced Agent Skills in October 2025 and published the format as an open standard that December. The Agent Skills site lists more than 40 agent products that support it. For Claude Code's own folders and extra fields, see Claude Code skills.

In Flywrite

Flywrite has no skills feature of its own. Skills are plain files, so .agents/skills/ and .claude/skills/ show in the file tree with the rest of your project. SKILL.md opens like any draft, and its front matter shows as a card where you edit the name and description in place. Start Claude Code, Codex or Cursor Agent in Flywrite's terminal with its launcher, and each reads skills from its own folders, as it does anywhere. More in Writing with Claude Code.

Questions

What's the difference between Agent Skills and MCP?

A skill teaches your agent how to do a task, in instructions it reads when it needs them. MCP connects your agent to other apps and data, so it can look things up and act there. They work together, since a skill can spell out a workflow that uses a connected app.

Are Agent Skills the same as Claude Code skills?

Claude Code skills follow the Agent Skills format and live in .claude/skills, with a few Claude Code-only extras such as disable-model-invocation. A skill that sticks to the standard's fields works in other agents too, once it's in a folder they read.

Sources
  1. Agent Skills: Overview. Checked .
  2. Agent Skills: Specification. Checked .
  3. Agent Skills: Client Showcase. Checked .
  4. Anthropic: Equipping agents for the real world with Agent Skills. Checked .
  5. Claude Code: Extend Claude with skills. Checked .
  6. OpenAI: Build skills. Checked .
  7. Cursor docs: Agent Skills. Checked .
  8. Cursor docs: Using the CLI agent. Checked .
  9. GitHub Docs: About agent skills. Checked .
  10. AGENTS.md. Checked .
  11. Claude Code: How Claude remembers your project. Checked .
  12. Model Context Protocol docs: Introduction. Checked .
  13. Flywrite docs: Writing with Claude Code. Checked .

Write with your agent in a writing app.

Buy now