Skills

What Skills Are

Skills are folders containing instructions, scripts, and resources that the AI discovers and loads dynamically when relevant to a task. They encapsulate a specific capability — instructions, context, and output format bundled together — so you don’t have to re-explain the same task every time.

Skills are now an open standard and being adopted broadly across platforms. Think of them as upgraded prompts: they package a prompt with its context into something reusable, shareable, and automatically invocable.

Key Characteristics

• Encapsulates a specific capability — instructions, context, and output format bundled together
• Dynamically loaded or directly invoked — the AI discovers and loads skills automatically when relevant, or you invoke them with a slash command (/plugin-name:command)
• Reusable across conversations — write once, use every time the task comes up
• Shareable — skills can be distributed to others through plugins or file sharing
• Becoming an open standard — the skill format is being adopted across compatible platforms

When to Use a Skill

Use a skill when:

• You find yourself writing the same prompt repeatedly
• A workflow step is well-defined enough to package as a repeatable routine
• Consistency matters — the output should follow the same structure every time
• You want others to be able to run the same task with the same quality

A good rule of thumb: if you give an AI the same instructions more than three times, it’s time to package those instructions as a skill.

Skills vs. Agents

A skill is a routine — it does one thing well when invoked. An agent is autonomous — it decides what to do, which tools to use, and when to invoke skills. Think of skills as tools in a toolbox and agents as the person using the toolbox.

Anatomy of a Skill

A skill is a folder containing:

skill-name/
├── SKILL.md          # Instructions — what the skill does and how
└── references/       # Optional context files the skill needs
    ├── style-guide.md
    └── template.md

The SKILL.md file contains the instructions. The references/ folder holds any context the skill needs — style guides, templates, examples, or data.

On Claude Code, skills with user_invocable: true and a command: field in their frontmatter can be invoked directly as slash commands.

Platform Implementations

Agent Skills are an open standard — the same SKILL.md format works across platforms. Each platform reads skill files from its own directory, but the file format is identical.

Not every platform supports skills natively yet. This table shows you what to expect:

Platform	Skill Support	Skill Directory / How It Works	Notes
Claude Chat & Cowork	✅ Native	Customize → Skills (ZIP upload) or Customize → Plugins	One skill library for Claude Chat on the web, the Chat tab in the Claude app, and Cowork. Plugins (paid plans) install from a marketplace and their skills work in all three; plugin hooks and sub-agents run only in Cowork
Claude Code	✅ Native	.claude/skills/	Also installable via plugins (/plugin install)
M365 Copilot Cowork	✅ Native	Customize → Skills → Upload skill, or OneDrive Documents/Cowork/skills/	Generally available; auto-discovered each session; up to 50 skills, 1 MB per SKILL.md, 20 companion files per skill
ChatGPT	✅ Native	Plugins → Skills (ZIP upload) or Plugins → Add marketplace	Skills run in Work (@skill) and Codex ($skill); plugin skills work in Chat too. Upload on paid plans except Free/Go; plugins on any paid plan; Enterprise/Edu admins enable Skills
OpenAI Codex (CLI, Codex tab, IDE)	✅ Native	~/.agents/skills/ or .agents/skills/	Same SKILL.md format; shares the plugin directory with ChatGPT
Gemini Spark	✅ Native	Skills → Upload in Spark mode	Personal Google Accounts with Google AI Pro/Ultra (not work/school accounts); web, Mac app, and mobile; sidebar Switch to Spark; /skill to invoke
Gemini Enterprise	✅ Native	Skills → + → Upload skill	Standard, Plus, and Pay-as-you-go editions; @skill or /skill; not usable inside Gemini Enterprise Agents
Gemini CLI / Antigravity	✅ Native	.gemini/skills/ or .agents/skills/	Same SKILL.md format
Cursor	✅ Native	.cursor/skills/, .claude/skills/, .codex/skills/, or .agents/skills/	Reads from Claude Code and OpenAI Codex directories too — no need to move files
VS Code Copilot	✅ Native	.github/skills/ or .agents/skills/	Same SKILL.md format

Legend: ✅ Native = platform officially supports Agent Skills. Every platform in this table can load the same SKILL.md file; only the upload path differs. For a platform not listed, paste the SKILL.md contents into the prompt or the workspace instructions.

Cross-platform convention

.agents/skills/ is a shared convention recognized by Cursor, OpenAI Codex CLI, Gemini CLI, and VS Code Copilot. Place your skills there to make them available across multiple platforms from one location.

Where Skills Run

The same SKILL.md format works everywhere, but the execution environment differs. Two factors matter:

1. Where the Skill executes — in a vendor-managed sandbox, or on your local machine 2. What it can access — only files you upload, or your full filesystem

Platform	Where Files Live	Execution Environment	Filesystem Access
Claude Chat (web and app)	Cloud (uploaded ZIP or installed plugin)	Cloud sandbox	Uploaded files only
Claude Cowork	Cloud (same library as Chat)	Cloud sandbox	The folder you connect
Claude Code (Desktop + Terminal)	Local (.claude/skills/)	Your machine	Full local filesystem
ChatGPT (Work)	Cloud (uploaded skill or installed plugin)	Cloud sandbox	Uploaded files and connected apps
Gemini Spark / Gemini Enterprise	Cloud (uploaded skill)	Cloud sandbox	Uploaded files and connected apps
Cursor	Local (.cursor/skills/ or .agents/skills/)	Your machine	Full local filesystem
OpenAI Codex (CLI, Codex tab, IDE)	Local (~/.agents/skills/ or .agents/skills/)	Your machine	Full local filesystem
Gemini CLI	Local (.gemini/skills/ or .agents/skills/)	Your machine	Full local filesystem
VS Code Copilot	Local (.github/skills/ or .agents/skills/)	Your machine	Full local filesystem
M365 Copilot Cowork	OneDrive (Documents/Cowork/skills/)	Microsoft cloud	OneDrive only

What this means for what you can build:

• Cloud-execution Skills (Claude Chat and Cowork, ChatGPT, Gemini Spark and Gemini Enterprise, M365 Copilot Cowork) excel at document creation, research, analysis, and anything where the AI works on content you provide in the conversation. They can’t read your local files, run commands on your machine, or touch your local development environment. (Cowork is the partial exception: it works inside the folder you connect, but still runs in Anthropic’s sandbox.)

• Local-execution Skills (Claude Code, Cursor, OpenAI Codex, Gemini CLI, VS Code Copilot) can read and write your actual project files, run terminal commands, integrate with local tools, and operate on your filesystem. Required for any workflow that needs to touch files outside the chat window.

The pragmatic default:

If you want to…	Build Skills for…
Create docs, slides, reports, research output	Claude Chat or Cowork, ChatGPT Work, Gemini Spark
Work on code or files in a project	Claude Code, Cursor, or OpenAI Codex
Operate inside Microsoft 365	M365 Copilot Cowork
Maximize portability across local AI tools	.agents/skills/ (works in Cursor, OpenAI Codex, Gemini CLI, VS Code Copilot)

Cloud sandbox capabilities

Cloud-execution Skills (like Claude.ai with code execution enabled) can still run code, generate files, search the web, and produce downloadable outputs. The sandbox is isolated from your local machine — files you upload go in, files Claude generates come out, but your local filesystem is untouched.

For full per-platform setup instructions, see the AI Workflow Framework Skills setup guide.

How to Install Skills

Each platform has its own install path. Pick yours — these pages link out to the official vendor docs for step-by-step instructions:

Platform	Skills setup
Claude (Chat, Cowork, Code)	Installing Skills on Claude
M365 Copilot Cowork	Skills on M365 Copilot
OpenAI (ChatGPT + Codex)	Skills on OpenAI
Google Gemini (Spark, Enterprise, CLI, Antigravity)	Skills on Google Gemini
Cursor	Skills on Cursor

VS Code Copilot and other editors: Drop the skill folder into .github/skills/. Use .agents/skills/ to share skills across Cursor, OpenAI Codex CLI, Gemini CLI, and VS Code Copilot from one location.

Setting up AI Workflow Framework skills?

See the framework setup guide — step-by-step for the seven AI Workflow Framework skills on each platform.

Skill, Workspace, or Prompt?

Approach	Best For	Example
Prompt	One-off or infrequent tasks	”Summarize this PDF”
Workspace	Recurring context without rigid steps	Client research workspace
Skill	Repeatable process with consistent format and standards	Weekly status report generation

What “workspace” means in your tool

Platforms name this differently — Claude calls them Projects, and other tools have their own equivalents. The idea is the same: a container holding standing instructions and reference files across many conversations, without prescribing a fixed procedure.

Guides

Guide	Description
How to Discover Your Best Agent Skills	Guided process to identify your highest-value skill candidates
Find Your Skill Candidates (Quick Prompt)	Fast conversational version — the AI interviews you and ranks candidates
The Complete Guide to Building Skills for Claude	Anthropic’s official guide covering skill structure, best practices, and advanced patterns
Improving skill-creator: Test, measure, and refine Agent Skills	How to test, measure, and iteratively refine skills using the skill-creator workflow

Related

• Agentic Building Blocks — Skills in the context of all building blocks
• AI Use Cases — what teams build with skills, organized by six primitives
• Prompts — the foundation that skills build on
• Agents — autonomous systems that invoke skills as part of multi-step workflows
• Agents & Skills — pre-built skills you can download or install