If you use Claude for the same job again and again, you end up pasting the same instructions into every chat: how you write a proposal, how you check an invoice, what your report should look like. A skill is how you stop doing that. You write the instructions once, Claude keeps a short label for them, and it pulls in the full version when your request matches.

This guide explains what a skill contains, how Claude decides to use one, how skills differ from the features they get confused with, and shows 10 real examples. Every fact below was checked against Anthropic’s own pages on 28 Sep 2026, and each example links to the file so you can read it yourself.

What a skill looks like

A skill is a folder. The one required file is SKILL.md: a short block at the top that names the skill and says when to use it, then the instructions in plain text. Anything else, such as scripts or reference files, is optional. Here is a small one we wrote as an example.

---
name: monthly-client-invoice
description: Draft, check and send the monthly client invoices in Xero from timesheet hours. Use on the last working day of the month, or when I say "run the monthly invoices".
---

# Send the monthly client invoice

## Steps
1. Open Xero and list every active client to invoice.
2. Pull each client's hours from the timesheet.
3. Draft an invoice per client and check the hours and rate.
4. Show me the drafts and wait for my go-ahead.
5. Send each approved invoice and mark it sent in the tracker.

## Never
- Send an invoice without my approval first.

The two lines at the top matter most. The name identifies the skill. The description tells Claude what it does and when to use it, and it is what Claude reads to decide whether this skill fits your request. A vague description means the skill never fires. A specific one, with the phrases you actually use, means it does.

How Claude uses a skill

Skills load in stages, so you can have many of them without filling Claude’s working memory. Anthropic’s documentation describes three levels:

1. Claude always sees the name and description

   Roughly a hundred tokens per skill, which is a sentence or two. That is all it needs to know the skill exists.

2. The instructions load when the skill matches

   When your request fits a description, Claude reads the body of SKILL.md.

3. Extra files load only if they are needed

   Reference documents and scripts stay out of the way until the steps call for them. Scripts run, and only their output comes back into the conversation.

Source: Anthropic’s Agent Skills overview.

Skills vs MCP, Projects, custom instructions and prompts

Skills are easy to confuse with the features around them. Here is how Anthropic separates them.

The useful way to hold it: MCP gives Claude access, a skill gives Claude the method. Connect Claude to your accounting system with MCP, and a skill tells it how you want the month-end invoices done. In Claude Code, skills have also merged with custom slash commands, so a command file and a skill folder can both create the same /command. Sources: Anthropic’s help centre and the Claude Code docs.

Where skills live

• Claude.ai: package the skill folder as a ZIP, then go to Customize, then Skills, click the plus, choose Create skill, then Upload a skill. Code execution has to be switched on. The folder name has to match the skill name.
• Claude Code: put the folder in ~/.claude/skills/ for yourself or .claude/skills/ inside a project.
• The API: upload skills through the skills endpoints, with the code execution tool enabled.

One thing catches people out: custom skills do not sync between those three places. A skill you upload to Claude.ai will not appear in Claude Code, and the other way round. Source: Anthropic’s guide to using skills.

Claude skill examples from Anthropic

Anthropic publishes example skills in a public repository. Six that are useful for everyday business work are below. Each card links to the file, so you can see exactly how it is written. Notice how much of each description is about when to use it.

skill-creatorPublished by Anthropic

Helps you create new skills, improve existing ones and test how well they perform.

Claude uses it: When you want to build, edit or tune a skill.

Source: SKILL.md on GitHub, checked 28 Sep 2026.

internal-commsPublished by Anthropic

Writes internal communications in the formats a company likes to use: status reports, leadership updates, newsletters, FAQs and incident reports.

Claude uses it: Whenever you ask for an internal communication.

Source: SKILL.md on GitHub, checked 28 Sep 2026.

brand-guidelinesPublished by Anthropic

Applies Anthropic's brand colours and typography to documents and other outputs.

Claude uses it: When brand or style guidelines apply to what Claude is making.

Source: SKILL.md on GitHub, checked 28 Sep 2026.

doc-coauthoringPublished by Anthropic

Walks you through writing proposals, specs and decision documents in three stages: gather context, refine the draft, then test it on a fresh reader.

Claude uses it: When you want to write documentation, a proposal or a spec.

Source: SKILL.md on GitHub, checked 28 Sep 2026.

docxPublished by Anthropic

Creates and edits Word documents, including headings, tables of contents, page numbers and tracked changes.

Claude uses it: Any time a Word document is the job. Its description also says when not to use it: not for PDFs, spreadsheets or Google Docs.

Source: SKILL.md on GitHub, checked 28 Sep 2026. Source-available for reference, not open source.

xlsxPublished by Anthropic

Creates, edits and cleans up spreadsheets, including formulas, formatting and charts.

Claude uses it: When a spreadsheet file is the main input or output.

Source: SKILL.md on GitHub, checked 28 Sep 2026. Source-available for reference, not open source.

Claude skill examples from other companies

Skills are not only an Anthropic thing. These are published by the companies themselves. They are all aimed at developers, which is where skills started, and we did not find a verified example from a non-technical business to include. That is a gap, and it is the reason the tool on this page starts from an ordinary SOP.

deploy-to-vercelPublished by Vercel

Deploys apps and websites to Vercel.

Claude uses it: When you say things like “deploy my app” or “push this live”.

Source: SKILL.md on GitHub, checked 28 Sep 2026.

code-reviewPublished by Sentry

Reviews pull requests following Sentry's own engineering practices, covering security, performance, testing and design.

Claude uses it: When you ask for a code review or feedback on a change.

Source: SKILL.md on GitHub, checked 28 Sep 2026.

durable-objectsPublished by Cloudflare

Helps build, debug and review code that uses Cloudflare Durable Objects.

Claude uses it: When the task involves Durable Objects.

Source: SKILL.md on GitHub, checked 28 Sep 2026.

stripe-best-practicesPublished by Stripe

Guides decisions when building a Stripe integration, such as which API to use.

Claude uses it: When you are planning or building a Stripe integration.

Source: SKILL.md on GitHub, checked 28 Sep 2026.

Logos are trademarks of their owners and are shown only to identify each publisher. Repositories change, so check the linked file for the current version.

What the good ones have in common

• The description says what it does and when to use it. Anthropic’s guidance is to write it in the third person and include the words a person would actually use. “Helps with documents” is too vague to trigger anything.
• One skill, one job. Every example above does one thing. The help centre’s advice is to keep one workflow per skill and start simple.
• They say when not to use it. The docx skill explicitly rules out PDFs, spreadsheets and Google Docs, so it does not fire when the job is something else.
• They stay short. Anthropic recommends keeping the SKILL.md body under 500 lines and moving detail into separate files.
• They include examples. A concrete input and a concrete output teach Claude more than a paragraph of explanation.

A note on description length, because Anthropic’s own pages disagree. The help-centre page on creating custom skills gives a 200-character limit. The open specification allows up to 1,024. Anthropic’s own published skills mostly run longer than 200: the six above range from about 240 to about 950 characters. If you are uploading to Claude.ai, keep yours under 200 to be safe.

Source: Anthropic’s skill authoring best practices.

How to create a Claude skill

1. Start from a job you already do

   Pick something you repeat that has clear steps: an invoice run, a client onboarding, a weekly report. If you have an SOP for it, you are most of the way there.

2. Write down what a person would just know

   What starts the job, what “done” looks like, what to do when something is missing, and which steps need your approval. These are the gaps that make a skill misfire.

3. Write a specific description

   Say what the skill does and the phrases that should trigger it.

4. Save it, package it and upload it

   Put it in SKILL.md inside a folder with the same name as the skill, zip the folder, and add it in Claude.

5. Run it on something real and tighten it

   Check in Claude’s reasoning that it actually read the skill, then fix whatever it got wrong.

You do not have to write the file by hand. Claude can draft a skill if you describe the task, and Anthropic publishes a skill-creator skill for that. If you already have a written procedure, the free tool on this site takes your SOP, asks about the gaps an AI would trip on, and gives you a finished SKILL.md with a description under 200 characters. It takes a couple of minutes.

Common mistakes and myths

• “Skills only exist in Claude Code.” They also run in Claude.ai and through the API, and other tools now support the open format.
• “Upload it once and it works everywhere.” Custom skills do not sync between Claude.ai, the API and Claude Code.
• “Anthropic’s document skills are open source.” The docx, pdf, pptx and xlsx skills are source-available for reference. Other skills in the repo are mostly Apache 2.0.
• “A skill is the same as custom instructions.” Custom instructions are always on. A skill loads only when it is relevant.
• “Everyone in the team gets my skill.” Skills you upload to Claude.ai are yours alone unless an owner on a Team or Enterprise plan provisions them.
• Trusting any skill you download. Read it first. A skill can tell Claude to run code.

Frequently asked questions