Agent skills are reusable workflows that AI tools follow to do a specific task the same way every time. A skill is a folder with a file called SKILL.md that holds the steps, the output format and any templates, examples or scripts the task needs. The AI loads the skill when the task comes up, so nobody has to paste the same instructions into every chat.
Skills are now an open standard. Anthropic created the format, and Claude, ChatGPT, Microsoft Copilot, Cursor, Gemini CLI and dozens of other AI tools use it. A skill written once can work across the AI tools your team already uses.
This guide covers what agent skills are, what businesses use them for, where they work and how to run the same skills across more than one AI tool. We checked the details against Anthropic, OpenAI, Microsoft and the Agent Skills standard on 2 October 2026.
The Short Answer
- A skill is a written method your AI follows. Think of it as a standard operating procedure the AI reads before it starts the work.
- It's a folder with a SKILL.md file. The file has a name, a description of when to use it and the instructions. Templates, examples and scripts can sit alongside.
- Businesses use skills for four things: standard operating procedures, processes and automations, consistent output and the functional work that runs the business.
- Skills are an open standard. Claude, ChatGPT, Microsoft Copilot, Cursor and Gemini CLI all read the same format.
- Each AI tool keeps its own copy. If your team uses more than one tool, keep one master version of each skill and update every copy from it.
What Are Agent Skills?
An agent skill packages up how a task should be done. It turns know-how that lives in someone's head, or in a document nobody opens, into instructions an AI tool can follow reliably.
Every skill has three parts:
- A name and description. The description tells the AI when the skill applies, so it's the most important line in the file.
- Instructions. The steps in order, what the skill needs, the output it should produce and the checks to run before it finishes.
- Resources. Anything the task relies on, such as a template, a worked example, a style guide or a script for work that has to be exact.
AI tools don't load every skill into every conversation. They read each skill's name and description first, then load the full instructions only when a task matches. That's why a team can have dozens of skills installed without slowing the AI down.
What Agent Skills Are Used For
Standard Operating Procedures A skill holds the steps for a task, so the AI follows the same procedure every time. Your SOPs become something every AI tool can run.
Processes And Automations Skills run multi-step work across your systems, such as updating deals in a CRM, preparing a board pack or reconciling accounts in Xero. Many AI tools can also run them on a schedule.
Consistent Output Every person's AI produces work in the same format, terminology and tone. Proposals, briefs and reports come out the same whoever runs them.
Functional Work Each skill has a job. A business builds up a set that covers how it runs:
| Area | Example skills |
|---|---|
| Sales | Prospect research, discovery call notes to next steps, proposals in your format |
| Delivery | Project set-up, status reports, meeting notes to actions |
| Finance | Monthly close commentary, invoicing checks, budget against actuals in plain language |
| Operations | SOP writer, weekly operations report, supplier comparisons |
| Marketing | Campaign briefs, content in your brand voice, performance summaries |
| People | Interview kits, role descriptions, policy questions answered from your handbook |
The Agent Skills Standard
Anthropic developed the skill format and released it as an open standard, Agent Skills, in December 2025. The specification is public, so any AI product can support it.
It has been widely adopted:
- OpenAI: ChatGPT and Codex
- Microsoft: GitHub Copilot, VS Code and Copilot Cowork, which accepts SKILL.md uploads
- Google: Gemini CLI
- Coding tools: Cursor and JetBrains
- Data platforms: Databricks and Snowflake
For a business, that means the time spent writing good skills isn't locked to one AI vendor. A skill built for Claude can usually be uploaded to ChatGPT or Copilot Cowork, and the other way round.
Where Agent Skills Work
| AI tool | Where skills live | How you add one | Sharing |
|---|---|---|---|
| Claude (chat, Cowork, Claude Code, Microsoft 365 add-ins) | Customize > Skills | Upload a ZIP, ask Claude to draft it or record it in Cowork | Team and Enterprise: share with people or groups, publish to the organisation or provision for everyone |
| ChatGPT (Business, Enterprise, Healthcare, Edu) and Codex | Plugins > Skills | Create with chat, build in the editor or upload | Share with people or groups, publish to the workspace or install for others |
| Microsoft Copilot Cowork | Customize > Skills, stored in your OneDrive | A guided conversation or upload a SKILL.md or ZIP | Share with specific people, and re-share to send updates |
| GitHub Copilot, VS Code, Cursor and Gemini CLI | Skill folders in a project or your user settings | Add a skill folder with a SKILL.md | Through the project's repository |
Skills, MCP, Projects and Prompts
These get confused, and they do different jobs:
- Skills teach the AI how to do a task. They load only when the task needs them.
- MCP connectors give the AI access to other systems and data, such as your CRM, Xero or SharePoint. A skill tells the AI how to use those connections well, so most useful set-ups use both.
- Projects hold background files and context that are always loaded for one piece of work.
- Prompts are one-off instructions typed into a chat. A skill is a prompt worth keeping, written properly and reused.
Skills Work Best on Your Company Context
A generic skill downloaded from the internet knows nothing about your business. The skills that make a difference are built on your company context:
- your terminology and the way you name things
- your templates and formats
- your tone of voice
- the knowledge base your people already maintain
The research backs this up. SkillsBench, a 2026 study across 18 AI model and agent set-ups, found curated skills lifted the average task pass rate from 33.9% to 50.5%. Focused skills beat large bundles. Skills an AI wrote for itself, with no person's knowledge behind them, did worse than no skills at all.
The lesson for a business: the value is in your expertise and your context. The AI does the writing, but the method has to come from the people who know the work.
Using Agent Skills Across More Than One AI Tool
Most teams don't run on one AI. Some people prefer Claude and some prefer ChatGPT. Many businesses also have Microsoft Copilot. The standard means the same skill can work in all of them, but each platform manages skills separately:
- Each tool keeps its own copy. Improve a skill in Claude and the copy in ChatGPT stays as it was.
- Sharing stays inside each platform. A skill shared in your ChatGPT workspace doesn't reach the people working in Claude.
- Connections differ. A skill that uses a connected system needs that connection set up in each tool.
- History is uneven. Some tools keep versions of a skill and some don't, so a bad edit can be hard to undo.
The fix is to treat skills as company assets:
- Keep one master version of each skill.
- Give each skill an owner.
- Track every change.
- Update every tool from the master.
That way your team can use whichever AI suits the work, and the business can change tools without starting again.
Keeping Agent Skills Safe
Skills can include scripts and can tell an AI to install software or connect to outside services. Treat them like any other software you bring into the business.
- Use skills from sources you trust. Read the files before turning on a skill you didn't write.
- Check scripts and links. Look closely at any code and any instruction that connects to an outside website.
- Turn on scanning where it exists. ChatGPT scans uploaded skills, and Claude Enterprise offers skill scanning.
The risk is measurable. In February 2026, security company Snyk scanned 3,984 skills on two public marketplaces and found 36.8% had at least one security flaw.
How ELab AI Helps
ELab AI is an AI advisory and implementation company. We work out where AI will make the biggest difference to how your business runs, then build the skills, connections and knowledge base to make it happen. We're platform-agnostic, so the skills we build work in Claude, ChatGPT and Copilot, and we maintain them as the tools change.
If you're choosing between tools, our guide to Copilot, ChatGPT, Claude and Grok compares prices, data storage and fit. Our AI enablement approach covers selection, set-up and ongoing management.
Getting Started
If you're working out where skills would make the biggest difference, take our AI Readiness Assessment. It looks at your people, process, knowledge and data, then technology.
If you'd like skills built and maintained across your business, tell us about your business or book a conversation.
