You found the wording that made an AI assistant produce exactly the summary you wanted. A few weeks later, you are typing a rougher version from memory and getting a weaker result.

An AI skill is a way to keep what worked. It is a saved, written method that an assistant can reuse for one kind of task. This is a different meaning from the skills people learn. Here, the skill belongs to the assistant, and you write it.

The idea is simple enough to start in an ordinary document. Some AI tools now have a dedicated feature for it.

A skill is a written method

Think about the note you would leave a capable new colleague for a task you do every week. It would not be one clever sentence. It would explain when the job comes up, what they need, the steps, what a good result looks like and how to check it.

A useful skill covers the same five things. In tools with a skills feature, the assistant reads it in two stages:

Inside a skill

Read in two stages

Read first

Name and description

Tells the assistant which requests this skill is for.

Your request matches, so it reads the rest

Then the full method
  1. WhenThe situation it is for, and when it is not.
  2. InputsThe notes, facts or examples it needs.
  3. MethodThe steps, plus anything to leave alone.
  4. ResultThe length, shape and tone of the finished work.
  5. CheckHow a person decides the result is acceptable.

Anthropic's explanation of Agent Skills calls this progressive disclosure. It is why a clear description matters as much as the method.

Next: Saved prompt, Gem or skill?

Saved prompt, Gem or skill?

Several features save your instructions. They differ in how much they hold and when they are used.

Four ways to save instructions

What each one holds, and who starts it

  • Saved prompt
    Holds
    One instruction.
    Starts
    You run it again.
    Microsoft 365 Copilot: save a prompt
  • Gem
    Holds
    Instructions, plus optional files.
    Starts
    You open it.
    Google Gemini: open a Gem
  • Custom instructions and projects
    Holds
    Background for every chat, or every chat in a project.
    Starts
    Always on.
  • Skill
    Holds
    A folder with a SKILL.md file: name, description, instructions, plus optional examples, templates or scripts.
    Starts
    The assistant loads it when a task calls for it.

Anthropic published the skill format as an open standard in December 2025. OpenAI says its ChatGPT skills follow that standard, so a skill made in one product can be installed in another.

At the time of writing, Claude's help centre lists Skills in Claude on its free and paid plans, with code execution turned on. OpenAI's help centre lists skills in ChatGPT for its Business, Enterprise, Healthcare and Edu plans. Features and plans change, so check the account you actually use.

Next: A worked example: site notes into a summary

A worked example: site notes into a summary

Imagine a fictional landscaping business. After each site visit, the owner types quick notes and asks an assistant to turn them into a short summary for the customer. The best results came when the instructions said to use only the notes and to flag anything missing.

Written as a skill, that method could look like this:

---
name: site-visit-summary
description: Turns a landscaper's site-visit notes into a short summary for the customer. Use when visit notes are supplied and a customer summary is requested.
---

Use only the notes supplied. Do not add prices, dates or materials
that are not in the notes.

1. List the work discussed, one plain line per item.
2. List anything the notes say is not included.
3. Put open questions for the owner in a separate list, after the
   customer text.
4. Keep the customer text under 150 words, warm and clear.

Before finishing, check that every item in the notes appears once
and that the summary contains nothing the notes do not support.

You do not need the file format to begin. The same text works pasted at the start of a conversation, or saved as a Gem or a project instruction. A dedicated skill becomes useful when you want the assistant to pick up the method by itself, or to share it with your team.

Notice what is not in it. The skill holds the method, not the business's prices or customer details. Facts that change with each job arrive with the job.

Next: Try it: write one and test it three ways

Try it: write one and test it three ways

Choose a task you have explained to an assistant at least three times. Write its five parts in a document, then test it with fictional or de-identified notes.

Three tests

Try each one with fictional notes

  1. Test 1 Normal notesCompare the result with the notes, line by line.Pass: Every item kept, nothing added.
  2. Test 2 A missing detailRemove one important fact, such as the garden bed measurements.Pass: It asks, instead of quietly filling the gap.
  3. Test 3 The wrong taskAsk for something unrelated, such as a social media post.Pass: The skill stays out of the way.

Fix the instruction behind each problem, not just the output. Then save the skill with a date and one line about what changed. Keep the previous version until the new one has proved itself.

Next: How to tell a skill is working

How to tell a skill is working

A good skill makes the result easier to check. It does not make checking unnecessary.

Before you rely on it

Four signs a skill is working

  • Keeps every supplied fact and adds none.
  • Asks when something important is missing.
  • Is used when it should be, and not when it should not.
  • Has a named owner and a date, so it is easy to trust and update.

If Claude skips your skill, its help centre suggests revising the description so it says more clearly when to use it.

Reviewing the output is still part of the job. How to check an AI answer walks through that review.

Next: Keep private details and unknown code out

Keep private details and unknown code out

A skill is a set of shared instructions, so treat it like a document other people may read. Keep passwords, account keys and customer records out of it. Claude's help centre warns against putting sensitive information such as API keys or passwords in a skill.

Australia's privacy regulator, the OAIC, recommends not entering personal information, particularly sensitive information, into publicly available generative AI tools. Use fictional or de-identified examples while you build and test.

Be careful with skills written by someone else. A skill can include scripts and instructions to connect to other services. Anthropic advises installing skills only from trusted sources, and OpenAI says its own scanning should not replace your review. If you cannot tell what a skill does, do not install it.

Next: Skills and connections work together

Skills and connections work together

Claude's help centre describes a skill and a connection such as MCP working together: MCP connects to services and data, while skills supply the knowledge of how to use them.

Two pieces, one job

The method and the route

SkillHow to do the task.
Connection, such as MCPA route to your tools and information.
TogetherThe assistant does the job. A person checks it.

Start with the method. If you are still choosing a first task, begin with where to start with AI. Once a method keeps working, writing it down is what makes it yours to reuse.