Claude Skills: How to Create, Install and Use Your Skills

You keep giving Claude the same instructions from one conversation to the next: your meeting notes format, your audit checklist, your style guide. A skill solves that problem. You write the method once, in a folder, and Claude finds it on its own whenever a request calls for it.

I start with the definition and structure of a skill, then installation and step-by-step creation, with a SKILL.md file generator. After that come the differences with projects, MCP and subagents, ten examples, six of them for SEO, and the security precautions.

What is a Claude skill?

A skill is a folder that brings together instructions, scripts and resources for a specific task. Anthropic's help center describes these folders as resources "that Claude loads dynamically": Claude spots the skills that are useful for the current request and reads only what it needs.

Anthropic launched skills on October 16, 2025, in the Claude app, in Claude Code and in the API. On December 18, 2025, the company added organization-wide skill management and a directory of skills built by partners, and published the format as an open standard called Agent Skills, documented at agentskills.io. So you'll come across both terms: "Claude Skills" refers to the feature in Claude products, "Agent Skills" to the file format.

In practice, a skill passes on know-how: how you write meeting notes, how you audit tags, how your brand writes a number. Claude already knows how to write, analyze and code. The skill teaches it your way of doing it.

How Claude picks and loads a skill

The mechanism is called progressive disclosure. Claude doesn't read all your skills in full with every message: it works in three levels, described in Anthropic's documentation.

Two practical consequences. You can install lots of skills without cluttering the conversation, since only their description stays loaded. And everything hinges on that description: if it doesn't say clearly when to use the skill, Claude won't open it.

The structure of a skill: SKILL.md, metadata and bundled files

A skill is a folder. Only the SKILL.md file is mandatory; the rest depends on the task. The Agent Skills specification recommends this layout:

seo-tag-audit/
├── SKILL.md       required: metadata and instructions
├── references/    optional: documentation to consult
├── scripts/       optional: code to run
└── assets/        optional: templates, layouts, data

The metadata: name and description

The file starts with a YAML block framed by two lines of three hyphens. Two fields are required:

• name: 64 characters maximum, lowercase letters, digits and hyphens only, with no hyphen at the start, at the end or doubled. It must match the folder name. The words "claude" and "anthropic" are reserved.
• description: what the skill does and when to use it, in 1,024 characters at most according to the specification. For an upload to the Claude app, the help center sets the limit at 200 characters: keep it short.

Anthropic recommends writing the description in the third person ("Checks the tags…" rather than "I can check…") and including the words you use in your requests. Compare:

• Weak description: "Helps with SEO."
• Good description: "Checks a site's title tags and meta descriptions and suggests fixes. Use for a tag audit, or a title or meta description audit."

The instructions in Markdown

Below the metadata, you write freely in Markdown: the steps, the expected format, the rules, an example of a good result. Three principles taken from the best practices Anthropic publishes:

• Don't explain what Claude already knows. It knows SEO and the CSV format. Write what is specific to you: your thresholds, your exceptions, your format.
• Stay under 500 lines. Beyond that, move the detail into reference files.
• Match the freedom to the risk. A fragile task, like a redirects file, calls for exact steps; an analysis task can handle general guidelines.

Bundled files and scripts

Reference files are read on demand: if the task doesn't need them, they cost nothing. Scripts run without being read, and only their output enters the conversation, which makes them more reliable and cheaper than code rewritten every time. Two rules: link to each file directly from SKILL.md, with no chain of references, and write paths with forward slashes, like references/thresholds.md.

Settings specific to Claude Code

Claude Code accepts extra fields in the metadata. The most useful ones:

• disable-model-invocation: true: the skill only runs manually, with /skill-name. Keep this for sensitive actions, like a deployment.
• user-invocable: false: the skill disappears from the menu and stays a reference that Claude consults on its own.
• allowed-tools: the tools Claude can use without asking for permission during the turn in which the skill runs.
• context: fork: the skill runs in an isolated subagent, which doesn't see the conversation history.

Watch out: for an upload to the app or through the API, only name, description, license, compatibility, metadata and allowed-tools are accepted. Any other field makes the upload fail.

Where Claude Skills work: the app, Claude Code and the API

The same folder format works everywhere, but each surface has its own rules.

Three points to remember. Code execution must be turned on, otherwise skills don't work. According to Anthropic's help center, skills work on every plan, from Free to Enterprise. And they don't sync: a skill uploaded to the app doesn't exist in the API or in Claude Code, so you have to install it on each surface.

Skills also work in Claude Cowork, Claude's agent mode for office work, where they often come inside plugins, alongside connectors and subagents.

How to install a Claude skill

In the Claude app

1. Open Settings > Capabilities and check that code execution and file creation is turned on.
2. Go to Customize > Skills.
3. Click +, then Create skill, then Upload a skill.
4. Choose the ZIP file. The skill folder must sit at the root of the ZIP: my-skill.zip contains my-skill/SKILL.md, not a standalone SKILL.md file.
5. Turn the skill on, then test it in a new conversation.

On the Team or Enterprise plan, the admin must first allow code execution and skills in Organization settings > Plugins and skills. In Cowork, the skills directory, opened from Customize then the + button, also offers ready-to-install skills, including partner skills from Notion, Figma, Canva or Atlassian.

In Claude Code

Claude Code reads skills straight from folders, with no upload. If the tool isn't on your machine yet, start by learning how to install Claude Code.

1. Copy the skill folder into ~/.claude/skills/ to have it in all your projects, or into .claude/skills/ at the root of a project to share it with your team through Git.
2. Type /skills to check that it shows up.
3. Let Claude use it on its own when the request matches, or run it with /skill-name.

To install the skills Anthropic publishes, add their repository as a plugin marketplace, then install the bundle you're interested in:

/plugin marketplace add anthropics/skills
/plugin install document-skills@anthropic-agent-skills
/plugin install example-skills@anthropic-agent-skills

Old custom commands, stored in .claude/commands/, keep working: Claude Code now treats them as skills.

Through the Claude API

On the developer side, skills go through the code execution tool. You declare the skill in the request's container parameter, with its ID: pptx, xlsx, docx or pdf for Anthropic's skills, or the ID of a skill you uploaded through the /v1/skills endpoints. The skill then runs in a container with no internet access, shared with the whole workspace.

Where to find skills for Claude

• The official repository anthropics/skills on GitHub: document, creative, development and communication skills. Anthropic states that they are provided for demonstration purposes and should be tested before you rely on them.
• The directory built into Claude, with Anthropic's skills and partner skills.
• Community catalogs: there are several, listing thousands of skills of very uneven quality. Read every file before installing; the security section below explains what to check.
• Your own skills, often the most useful, because they contain what no catalog knows: your method.

How to create a Claude skill step by step

A simple skill fits in a text file. The method below applies to every surface; only the last step changes.

1. Pick a task you explain often. The right candidate is the instruction you copy from one conversation to the next: a report format, a checklist, a style guide.
2. Do the task once with Claude, without a skill. Note what you have to clarify or correct: that's exactly the information Claude is missing, so it's the content of the skill.
3. Create the folder and the file. A folder named after the skill, lowercase with hyphens, and inside it a SKILL.md file.
4. Write the metadata. The name, then a description that says what the skill does and when to use it, with the words from your requests.
5. Write the instructions. The steps in order, the output format, the rules to follow and, if possible, an example of a successful result.
6. Add the useful files. A scoring grid in references/, a template in assets/, a check script in scripts/. None of it is required.
7. Test it on three real requests. In a new conversation, phrase the request as usual, without mentioning the skill. Claude's displayed reasoning should show that it reads the skill. If it ignores it, rework the description.
8. Install it for your surface: ZIP in the app, folder in Claude Code, upload through the API.

Without writing a line: ask Claude to create the skill

Claude knows the format. After doing a task with it, ask: "Create a skill that captures the method we just applied, with the rules I gave you." Anthropic also publishes a dedicated skill, skill-creator, that guides you through creating, editing and testing a skill. Always reread the result: a model tends to add unnecessary explanations, which you should cut.

In Claude Cowork on Mac, the Pro, Max and Team plans also let you record a skill: you perform the task on screen, for ten minutes at most, and Claude proposes a skill based on the recording. The feature isn't available in chat, on Windows, or on the Free and Enterprise plans.

A complete SKILL.md example

Here is a complete SEO skill, ready to adapt. It mirrors the tag-checking pipeline I run on client projects with Claude Code.

---
name: seo-tag-audit
description: "Checks a site's title tags and meta descriptions and suggests fixes. Use for a tag audit, or a title or meta description audit."
---

# SEO tag audit

## Steps
1. Read the crawl export provided (CSV) or crawl the URLs given.
2. Find titles that are missing, duplicated, under 30 or over 60 characters.
3. Do the same for meta descriptions (missing, duplicated, over 155 characters).
4. Suggest a rewrite for each faulty page, with the character count.
5. Don't change any file without explicit approval.

## Output format
A table: URL, issue, current tag, suggestion, character count.
Then the five pages to fix first, ranked by traffic if the export includes it.

## Rules
- The title contains the query the page targets, at the start if possible.
- No unverifiable claims ("the best", "#1").
- Flag ambiguous cases in a list for review instead of deciding.

## Reference
Site-specific thresholds and exceptions: see references/thresholds.md

To write your own, the generator below builds the file from your answers, checks the name and the description length, and tells you where to put it for the surface you chose.

Skills, projects, MCP, subagents: what's the difference?

Claude offers several ways to give it context, and they're easy to mix up. Anthropic sums up the distinction like this: a project tells Claude what it needs to know, a skill tells it how to do things, and MCP connects it to your data.

Skill or project. A project stores material (documents, general instructions) available in all its conversations. A skill stores a procedure, loaded only when it's needed, in any conversation, whatever the project.

Skill or MCP. MCP gives Claude tools: reading Search Console, launching a crawl. The skill tells it what to do with them: which pages to compare, which thresholds to apply, which table to return. The two combine, and that's often the best setup. I explain how an MCP server works in a dedicated article.

Skill or subagent. A subagent is a full assistant, with its own context, that the main agent delegates part of the work to. A skill is know-how that any instance of Claude can load. In Claude Code, the two meet with context: fork, which runs a skill inside a subagent. For agent logic in general, see how to build an AI agent.

10 Claude skills examples, including 6 for SEO

The first four are provided by Anthropic or its partners, or take a few lines to write. The next six cover the SEO tasks I automate with Claude Code for clients: each has its own input data, an explicit rule and a verifiable result, which makes them good candidates.

Ready-made skills and business skills

1. Anthropic's document skills

Word, Excel, PowerPoint and PDF: Claude uses them on its own when you ask for a document, in the app as through the API. They are what let it produce a spreadsheet with formulas or a properly laid-out presentation.

2. skill-creator, the skill that creates skills

It asks the right questions, writes the SKILL.md, helps you edit it and measure whether it triggers as intended. It's in Anthropic's official repository.

3. Partner skills

Notion, Figma, Canva and Atlassian publish skills that teach Claude to work with their tools. You install them from Claude's skills directory, like Anthropic's.

4. A brand style guide skill

Your tone, your banned words, how to write numbers, two reference texts in references/. It's a good first skill for a team, because a style guide applies to everything that gets written.

---
name: brand-style-guide
description: "Applies the company's style guide to any text written or reviewed. Use for an email, a post, a web page or a proofread."
---

6 SEO skills to create

For these skills, Claude needs access to the data: the site's files, a crawler, Search Console. That access goes through MCP connectors, like the Screaming Frog MCP for crawls. The skill carries the method. Other ways to use Claude for search are covered in Claude for SEO.

5. Title and meta tag audit

The obvious starting point: the task is mechanical, can be checked at a glance and becomes impossible to keep up to date by hand across several hundred pages. The complete example is above. The guardrail: Claude suggests, you review, because a tag is what visitors see in search results.

6. Internal linking audit

Extract the link structure, find orphan pages and pages buried too deep, suggest three inbound links per priority page with their anchor text. The rule I enforce in my instructions: no generic anchors, every anchor must describe the target page.

---
name: internal-linking-audit
description: "Analyzes internal linking from a crawl, finds orphan or overly deep pages and suggests links with their anchor text. Use for an internal links audit."
---

7. JSON-LD structured data

Generate the markup for a catalog or a set of articles, then validate the syntax. Absolute rule: the markup reflects what is visible on the page, without filling in missing information with a plausible value.

---
name: structured-data
description: "Generates a page's JSON-LD markup from its visible content, without inventing missing data. Use to add or fix structured data."
---

8. Migration redirect plan

Match each old URL to its equivalent, flag addresses with no obvious match instead of deciding, generate the file while avoiding chains and loops. This skill is better kept manual: in Claude Code, add the line below.

---
name: redirect-plan
description: "Builds the 301 redirect plan for a migration, one old URL to one specific destination. Use during a redesign or a URL change."
disable-model-invocation: true
---

9. Monthly Search Console report

Compare two periods, isolate the drops, cross-reference them with changes made in between, wrap up with three actions. A methodological precaution to write into the skill: compare periods of at least 28 days, since a one-week change means nothing.

---
name: search-console-report
description: "Compares two periods of Search Console data, isolates declining pages and queries and suggests three actions. Use for a monthly SEO tracking report."
---

10. Pre-publication check

Tag length, heading hierarchy, valid internal links, alt attributes, unsourced claims. A systematic technical review that doesn't replace the editorial one: a text can pass every check and still be bad.

---
name: pre-publish-check
description: "Checks a page before it goes live: tags, headings, internal links, alt attributes and unsourced claims. Use before publishing an article or a page."
---

Claude Skills limits and security

A skill isn't just text: it can contain scripts and steer the tools Claude uses. Anthropic recommends treating it like installing software.

• Only install skills you trust, your own or Anthropic's. A malicious skill can push Claude to run code or send data outside.
• Read everything before installing a skill found online: SKILL.md, scripts, resources. Be wary of unexpected network calls and of skills that fetch instructions from an external URL, whose content can change after you've checked it.
• No syncing between the app, Claude Code and the API: each surface has its own copy, which you have to keep up to date.
• Triggering depends on the description. Too vague, and Claude ignores the skill; too broad, and it loads it by mistake. Two skills with similar descriptions get in each other's way.
• Every skill takes up a little context permanently, about 100 tokens of metadata: fifty skills already add up to some 5,000 tokens.
• On the API side, there's no internet access and no package installation, and skills are not covered by zero data retention agreements.
• In the enterprise, Enterprise organizations can turn on scanning of the content of skills uploaded to Claude and Cowork.

If you want business skills designed, tested on your own files and delivered with these guardrails, that's what I do in an AI automation engagement.

Written by

Anthony Courtin

SEO, GEO & AI automation consultant

I help SMEs and independent professionals with their visibility (Google, ChatGPT, Perplexity) and with the real-world use of AI at work. I run on-site sessions to train teams and rethink their processes. Everything I write here, I use in client work.

Follow me on LinkedIn My YouTube channel (in French)