The idea
Think of a job you do the same way every week. The Monday status update. Turning meeting notes into a list of actions. The first reply to a supplier who asks for a quote. Most people either type the instructions again each time, or keep a document of favourite prompts to copy in. That document is a prompt pack.
A skill does the same job, but your assistant keeps it and reaches for it by itself. It is a folder with one file in it, called SKILL.md. At the top of the file sit two lines: a name, and a description of what the skill does and when to use it. Below them come the steps, in plain words, the way you would explain the job to a new colleague. The folder can also hold an example of a good result, your template, or a short reference file.
The clever part is when it loads. The open Agent Skills standard describes three stages. At the start, the assistant loads "only the name and description of each available skill". When a task matches a description, it reads the full instructions. Then it follows them. So you can keep many skills on hand, and the description decides when each one is used. That is why the description must say plainly when to use it.
One file, three assistants
The format was, in the standard's own words, "originally developed by Anthropic, released as an open standard", and it is published at agentskills.io. As of Oct 2026, Claude, ChatGPT and Gemini all accept a SKILL.md file. Each tool still has its own limits (see below), so add the skill and test it in each tool you use.
Why the prompt pack is retiring
- OpenAI: custom GPTs are planned for retirement, with a move to plugins, and a plugin can hold skills alongside apps (release notes, 11 Sep 2026).
- Google: Gems become skills from November 2026 on personal accounts, and in 2027 on work and school accounts.
- Anthropic: Claude takes skills in Customize > Skills, and the format began there.
Read a skill before you add it
A skill is a set of instructions your assistant will follow, and it can include scripts. Anthropic's advice is to read a shared skill's files before you turn it on, and it says plainly that skills from others are not reviewed by Anthropic. Google limits what an uploaded skill can do instead: no scripts that need the internet. The safe habit is the same in all three: add skills only from sources you trust, and read them first, even one a colleague shared.
A small example
This is an example, made up for this lesson. It is a whole skill. The folder is called weekly-team-update, and inside it is one file, SKILL.md:
---
name: weekly-team-update
description: Turns my rough notes into our one-page weekly team update. Use when I ask for the weekly update or paste notes for it. Drafts only.
---
# Weekly team update
1. Use only the notes I give you. Do not add facts.
2. Write three headings: Done this week, Next week, Needs a decision.
3. Keep each point to one line, with the owner's name in brackets.
4. If something is unclear or missing, list it under "Questions for you". Do not guess.
5. Draft only. Never send, post or share anything.
## What a person must check
- Every date, number and name against the notes.
Three habits are built in: the name is lowercase with hyphens (the standard and Google both require it), the description says when to use it, and the skill drafts and never sends.
Common mistakes
- A vague description. "Helps with reports" runs at the wrong moments or never. Say what it does and when to use it, in the words people actually type.
- One skill for everything. Make separate skills for separate jobs, so each description can say plainly when it applies.
- Private data inside the skill. Names, customer details, prices you have agreed, passwords. Skills get shared and synced. Keep the method in the skill and the data in its own system.
- A skill that acts on its own. Write "draft only" and "never send" into it. Sending, posting and paying stay with a person, or behind the tool's own approval.
- Adding a skill you have not read, from a directory, a website or a colleague.
- Assuming one test is enough. Names, limits and how you call a skill (/ or @) differ between the three tools. Test it in each tool you use.
- Waiting for the switch-off. If your team relies on a custom GPT or a Gem, plan the move now and test the replacement before the date your account shows.
What to check before you trust it
- Does it run at the right moment? Try three made-up requests: one it should handle, one close but different that it should leave alone, and one with information missing. It should run on the first only, and ask about the third.
- Did the right skill run? Call it by name to be sure: type / and pick it in Claude or Gemini, or type @ and pick it in ChatGPT.
- Does the result follow the steps? Right format, nothing added that was not in your input, questions asked instead of guesses.
- Is the content right? A skill makes the work consistent, not correct. Check facts, figures and names in what it produces, as you would any AI work.
- Is it still current? These tools change every month. Re-test your skills after a big update, and at least once a month.
Keep it safe
- Read every skill before you add it, and add only skills from sources you trust.
- No secrets, customer data or personal details in a skill file.
- Draft only: a skill never sends, posts, pays or deletes by itself.
- On a work account, follow your organisation's way of sharing skills, so the team uses one checked copy.