AGENT SKILLS

Agent Skills in Claude Code, Codex, Copilot & Cursor.

Agent Skills are an open standard — but how each tool implements them differs in folder location, loading behaviour, and distribution model. This guide covers the current state of Agent Skills support in the four major AI coding agents, plus other notable implementations. Support, folder paths, and exact behaviour evolve quickly — always verify against each tool's current documentation before relying on any detail here.

Important: this information changes

Agent Skills support is actively evolving. Folder paths, config keys, and feature availability differ between tools and change between releases. Use this guide as a starting point, then verify against each tool's current official documentation. Links to current docs are provided in each section.

Claude Code

Claude Code (Anthropic's CLI) was the first tool to implement the Agent Skills standard. Skills are loaded from the .agents/skills/ directory at the project root by default, and from ~/.claude/skills/ for user-level skills available across all projects.

your-project/
└── .agents/
    └── skills/
        ├── playwright-test-gen/
        │   └── SKILL.md
        └── api-test-generator/
            └── SKILL.md

Discovery: Claude Code scans all subdirectories of the skills folder at startup and reads the name and description fields from each SKILL.md.

Activation: When the user's message matches a skill description, the full SKILL.md body is loaded into context. Supporting files are loaded on demand during execution.

Slash command: You can also activate a skill explicitly via /skill <name> in the Claude Code CLI.

Check current docs

Folder paths and the slash command interface may change. Verify at the Claude Code documentation before pinning these paths in CI or shared tooling.

OpenAI Codex

OpenAI Codex reads skills from .agents/skills/ directories within a repository. User-level configuration is managed via ~/.codex/config.toml.

# ~/.codex/config.toml — user-level skill configuration
[skills]
paths = [
  "~/.codex/skills",          # global user skills
  ".agents/skills",           # project-level skills (relative to repo root)
]

Distribution beyond a repo: For distributing skills outside a single repository, Codex uses a plugins system — a packaged format that bundles skills for discovery and installation. If you want to share skills across many projects without manually copying directories, investigate the Codex plugin format.

Progressive disclosure: Codex implements the same three-stage model (Discovery → Activation → Execution) as the open standard, but its exact implementation details — particularly around multi-skill composition and execution sandbox — should be verified in the Codex documentation.

Check current docs

Codex's skills implementation and config.toml schema evolve rapidly. The paths and keys above were accurate at the time of writing; check the OpenAI Codex documentation for current values.

GitHub Copilot

GitHub Copilot loads Agent Skills from the .github/skills/ directory in a repository. This follows GitHub's existing convention of storing repository-level configuration under .github/.

your-project/
└── .github/
    └── skills/
        ├── playwright-test-gen/
        │   └── SKILL.md
        └── api-test-generator/
            └── SKILL.md

Repository context: Because skills live in .github/, they are automatically version-controlled alongside the code they support. Teams can review skill changes in pull requests and pin skills to specific commits.

Access control: Skills in a private repository are accessible only to users with repository access — useful for keeping company-specific testing standards internal.

Check current docs

GitHub Copilot's Agent Skills support was introduced as part of the broader agent-skills open standard adoption. Verify the current folder convention and any Copilot-specific configuration in the GitHub Copilot documentation.

VS Code (native Copilot integration)

VS Code includes native Copilot agent-skills support. In addition to loading skills from the repository's .github/skills/ directory, VS Code provides a built-in slash command:

/create-skill

Running /create-skill in the Copilot chat panel opens a guided flow to generate a new SKILL.md file. It prompts you for the skill's name, description, and key instructions, then writes the file to the appropriate location in your workspace.

Workspace vs. user scope: VS Code distinguishes between workspace-level skills (in the repository, shared with the team) and user-level skills (personal, not committed to the repo). Consult the VS Code Copilot settings for the user-level skills path.

Check current docs

The /create-skill command and workspace skill paths are subject to change with VS Code and Copilot extension updates. Check the VS Code Copilot documentation for the current state.

Cursor

Cursor supports SKILL.md for workspace-level customisation. Skills placed in the workspace's skills directory are picked up by the Cursor agent and used to inform its behaviour for matching tasks.

your-project/
└── .cursor/
    └── skills/
        └── playwright-test-gen/
            └── SKILL.md

Rules vs. Skills: Cursor also has a “Rules” system (.cursorrules / .cursor/rules/). Rules are always-on context injected into every conversation; Skills are activated selectively by the agent based on the task. Use Rules for global project conventions and Skills for specific, activatable workflows.

Check current docs

Cursor's skills folder path and the interaction between Rules and Skills may change. Verify against Cursor's documentation before relying on the paths above.

Other implementations

The open standard has been adopted beyond the four major tools:

Goose (Block)

Block's open-source Goose agent supports Agent Skills. Skills are loaded from a configurable path in the Goose config file. Check the Goose documentation for the current default location.

Spring AI

The Spring AI project includes an implementation of the Agent Skills standard for Java/Spring-based AI applications. Skills can be defined alongside Spring beans and loaded as part of the application context.

agentskills.io

The canonical spec lives at agentskills.io, coordinated by the Agentic AI Foundation under the Linux Foundation. The spec defines the mandatory fields, optional directories, and the progressive disclosure protocol. New implementations should conform to it.

Portability: write once, verify per tool

The open standard means a SKILL.md written for Claude Code should work in Codex, Copilot, and Cursor with no changes to the file content — only the folder location differs.

ToolProject-level folderUser-level folder
Claude Code.agents/skills/~/.claude/skills/
OpenAI Codex.agents/skills/~/.codex/skills/ (via config.toml)
GitHub Copilot.github/skills/—
VS Code (Copilot).github/skills/Check VS Code Copilot settings
Cursor.cursor/skills/—

Practical portability strategy: Keep your canonical skills in .agents/skills/ (works for Claude Code and Codex). Add symlinks or a copy step in your project setup script for tools that use a different path (.github/skills/ for Copilot, .cursor/skills/ for Cursor). This avoids duplicating skill content while maintaining compatibility.

Check current docs

All paths in the table above should be verified against current tool documentation before use in production tooling. The spec at agentskills.io is the authoritative reference for the standard; individual tools may deviate from or extend it.

See QA skill examples →How to build a skill