Claude skills: write them once, use them in every AI tool

In short

• Claude skills are folders with a SKILL.md file: a name, a description of when the skill applies, and the instructions themselves. That format, Agent Skills, is an open standard that ChatGPT, Microsoft Copilot and Gemini can also read.
• The catch: the tools do not sync with each other. Every upload is a copy, and the package differs per tool.
• So write each skill once as a document with an owner, keep that version in one place and install it in each tool from there.

A skill is a reusable way of working for AI: a folder with a SKILL.md file that says what the skill does, when it applies and which steps the AI follows. Anthropic started the format in Claude and on 22 October 2026 donated it as an open standard to the Agentic AI Foundation, part of the Linux Foundation. According to Anthropic, more than 70 platforms support it, including ChatGPT, Cursor, GitHub Copilot and VS Code (Anthropic). Google also builds skills in Gemini on SKILL.md.

For a team that keeps pasting the same instructions into a custom GPT, a Gem and a Claude Project, this means you can write a way of working down properly once and use it in every tool. The main condition is that you stay in charge of the version yourself. No tool picks up a change made in another.

What are Claude skills and agent skills?

Agent Skills is the name of the standard; Claude skills is what Anthropic calls them inside Claude. According to the specification, a skill is a folder containing at least a SKILL.md. At the top are two required fields: name (lowercase letters, numbers and hyphens, up to 64 characters, matching the folder name) and description (up to 1,024 characters, saying what the skill does and when to use it). Below that come the instructions in plain text. You can add files such as examples, templates or scripts.

What sets a skill apart from a saved prompt is how it loads. The AI first sees only the name and description of each skill. It reads the instructions only when your request matches a description (Claude documentation). That is why you can have many skills switched on at once, and why the description is the most important piece of text: it is the only part the AI reads before choosing.

Anthropic also sets out how this differs from Projects. A Project holds background knowledge that is loaded in every chat in that project. A skill is a procedure for one task that loads only when a request calls for it. For more on Claude itself, read What is Claude?

Which AI tools read a skill, and how do you install it?

What the vendors themselves say, checked on 27 October 2026:

Sources per row: Claude, OpenAI Academy, Microsoft Learn, GitHub Docs, Gemini Apps Help and Google Workspace Updates.

Two things stand out. The ZIP for Claude and the one for Copilot Cowork are built differently: Claude looks for folder-name/SKILL.md inside the archive, Cowork for SKILL.md at the top. One folder, two packages. And Google writes that skills in Gemini are built on the open SKILL.md standard and that you can copy in skills from other platforms (Google Workspace Updates). Google does not say whether Gemini follows the full Agent Skills specification, nor whether it runs scripts inside a skill. Assume only the instructions are certain to carry over.

How do you build a skill that works in every tool?

Write the skill as a document a colleague could follow without AI. Then it matters little which tool reads it. Five parts:

1. Purpose. What the skill produces and when to use it. This becomes the description. Use the words someone would use when asking for this task.
2. Input. What the AI needs to know before it starts, and what it does when something is missing.
3. Steps. The method, in a fixed order.
4. Check. What the AI reviews in its own output before handing it back.
5. Example. An approved result, so the AI can see what good looks like.

On top of that comes an owner: the person who maintains the skill, approves changes and retires it when it no longer holds. The standard has a free-form field for this, metadata, with text values.

In this fictional example, a marketing team records what a campaign brief looks like:

---
name: campaign-brief
description: Writes a one-page campaign brief in the marketing team's standard format. Use this skill whenever someone asks for a brief, briefing or outline for a campaign, promotion or launch.
metadata:
  owner: "Marketing manager"
  version: "1.2"
  updated: "2026-10-27"
---

# Campaign brief

## Purpose
A brief the agency and the team can start from without follow-up questions.

## Input
Ask for anything missing: product, audience, goal with a number,
budget, duration and channels. Do not invent figures. Mark any
missing item as "to be added".

## Steps
1. Summarise the goal in one sentence.
2. Describe the audience and their problem.
3. Write the core message with three proof points.
4. Put channels, timing and budget in a table.

## Check
- Is the goal measurable?
- Does every figure come from the input?
- Does the brief fit on one page?

## Example
Follow the structure of references/example-brief.md.

Keep a skill small and focused: one task per skill. Anthropic recommends keeping the main file under 500 lines and moving long background material into separate files. Instructions that should apply in every conversation, such as your tone of voice, belong in the standing instructions or the Project, not in a skill. Looking for tasks to start with? See AI prompts for marketers by role: any prompt you rewrite every week is a candidate.

Where do you keep skills, and who maintains them?

Keep the skill folder in one shared place you already use, such as SharePoint, Google Drive or a repository. That is the master version. Whatever sits in Claude, ChatGPT, Copilot or Gemini is a copy.

Every change follows the same route:

1. The owner edits the master and raises the version number.
2. They test the skill with a few real requests in a new conversation.
3. They install the new version in every tool that uses the skill. In Copilot Cowork, Re-share sends the latest version to everyone you shared it with.

Keep a short list next to the master of the tools the skill is installed in. Without it, an old version keeps running somewhere.

Only upload skills you can read through. A skill acts as instructions to the AI, and Anthropic does not review skills you upload yourself or receive from someone else, so read the SKILL.md and its files first (Claude documentation). Microsoft gives the same warning. Never put passwords or API keys in a skill: whoever gets the skill gets all its files.

What should you do with your custom GPTs, Gems and Projects?

Start with a list. Ask the team which GPTs, Gems and Projects they use, who made them and what for. Then decide per item:

• A recurring task with steps becomes a skill in the master folder, with an owner.
• Background knowledge that is always needed stays in a Project or the standing instructions.
• Anything nobody uses any more can go.

For Gems there is a date. Google is replacing Gems with skills. For business Workspace accounts, Gems can no longer be created or used from no sooner than 1 March 2027 (Google Workspace Updates). On 13 October Google said skills had started rolling out in the Gemini app, but that the rollout to Workspace Studio and the Workspace apps has been delayed. For personal accounts, Google's own sources disagree: Workspace Updates says Gems move to the Settings panel on 17 November and can still be used there, while Gemini Apps Help says they become read-only that day and are removed on 17 December. What to sort out per Gem is covered in Gemini Gems are becoming skills.

If your team uses several tools side by side, ChatGPT, Claude or Copilot for your team helps with the choice. With skills in a master folder, that choice no longer locks in your ways of working.

When is a skill the wrong choice?

A skill suits work that comes back often and has to be done the same way each time. For a one-off request, just type what you want. If you need live data from another system, you need a connector: a skill can describe how the AI should use a connector, but cannot reach a system itself.

Bear in mind, too, that the standard is young. The Agentic AI Foundation is still working on versioned releases, a formal change process and security guidance. Test a skill in every tool you install it in, and do not rely on a script if you do not know whether that tool runs it.

Sources

• Anthropic, Donating Agent Skills to the Agentic AI Foundation, 22 October 2026.
• Agent Skills, Specification, accessed 27 October 2026.
• Claude, Skills overview and Create custom skills, accessed 27 October 2026.
• Claude Help Center, Use skills in Claude, accessed 27 October 2026.
• OpenAI Academy, Skills, accessed 27 October 2026.
• Microsoft Learn, Customize Copilot Cowork, updated 21 October 2026.
• GitHub Docs, About agent skills, accessed 27 October 2026.
• Google Workspace Updates, Introducing skills in the Gemini app and Workspace, plus what's next for Gems, September 2026, updated 13 October 2026.
• Gemini Apps Help, About the transition from Gems to skills, accessed 27 October 2026.