What Are Claude Skills? A Marketer's Guide
Claude Skills are reusable folders of instructions and scripts Claude uses automatically when relevant. How Agent Skills work in Claude Code and claude.ai.

On this page 13 sections
- What are Claude Skills?
- How do Claude Skills work?
- How does Claude decide which skill to use?
- What is inside a skill folder?
- Which skills are available, and who makes them?
- How do skills compare with other Claude capabilities?
- What is the Agent Skills open standard?
- Where do skills run: claude.ai, Claude Code or the API?
- How do you install and use a skill?
- How do you create skills of your own?
- Which use cases and workflows suit a skill?
- How do you check that a skill is safe?
- Use a skill to keep product reviews consistent
Short answer: Claude Skills are folders of instructions, scripts and reference files that teach Claude how to do a specific task your way. Each skill has a SKILL.md file whose name and description tell Claude when it applies; the full instructions load only when a request matches. Skills run on claude.ai, Claude Code and the Claude API, and follow the Agent Skills open standard.
What are Claude Skills?
Claude Skills are folders of instructions, scripts and resources that Claude loads when a task calls for them. Each folder holds a SKILL.md file: a short header with a name and a description, then Markdown instructions Claude follows, optionally with scripts, templates and reference material beside it. Anthropic's own analogy is an onboarding guide for a new hire.
That is the right mental model. A skill teaches Claude a procedure rather than new powers: it is a file that tells Claude exactly how you do a specific task, so the result comes out the same way every time.
Anthropic introduced the format as Agent Skills on 16 October 2025 and published it as an open standard that December. Most people use AI assistants by retyping the same brief into each chat. Skills are reusable: the brief is written once, and Claude applies it automatically whenever it fits.
How do Claude Skills work?
Skills work by progressive disclosure: Claude reads a little about every skill up front and the rest only when it is needed. That is why many skills can sit installed without filling the context window, the working memory Claude has for one conversation: Claude loads what it needs and nothing else.
- Selection
Claude decides whether the skill is relevant.
- Instructions
The skill supplies the workflow and resources.
- Execution
Available tools and permissions still constrain actions.
A skill can describe a process without granting new service access.
How do skills load into Claude's context?
Skills load in three levels, and each costs context only at its own moment. The figures below are from the official Agent Skills documentation, so even a large skills directory costs little until a request uses one.
Level | When Claude reads it | Rough context cost | What it contains |
|---|---|---|---|
1. Metadata | At the start of a session, in the system prompt | About 100 tokens per skill | The name and description from the frontmatter |
2. Instructions | When a request matches the description | Under 5,000 tokens | The SKILL.md body: workflow, rules and examples |
3. Resources and scripts | Only when the instructions point to them | Nothing until opened | Reference files, templates and executable scripts |
When a skill bundles a script for Claude to run, Claude receives only the output, so the code itself never spends a token of context. That is how the document skills produce a real spreadsheet or Word doc rather than a description of one: the instructions say what to build and the script builds it.
How does Claude decide which skill to use?
Claude decides whether to load a skill by comparing your request with the description of every available skill, then uses the one that can match your request. Ask for a quarterly report as a spreadsheet and the Excel skill loads without being named. The description is the only part Claude sees before it chooses, so it has to say both what the skill does and when to use it.
You can also ask Claude for a skill by name. In Claude Code you type /skill-name to run it, and a skill whose frontmatter sets disable-model-invocation: true runs only when you call it that way.
When two skills cover the same use case, naming one is the reliable route. The first time Claude meets a vague description, it may simply skip the skill, which is why how to write a Claude Skill that loads spends most of its length on that one line.
What is inside a skill folder?
Each skill folder needs one file, SKILL.md; the other skill files are optional and hold whatever the task needs. The table shows a typical layout for a skill that reviews product pages.
File | Role | Required |
|---|---|---|
| Frontmatter with | Yes |
| Detail Claude opens only when the instructions say so, such as brand guidelines | No |
| An executable script for a step that must be exact | No |
| A fixed output format to fill in | No |
The Agent Skills rules for the header are short. The name is at most 64 characters of lowercase letters, numbers and hyphens, and on the Claude API it cannot contain the reserved words "anthropic" or "claude". The description must not be empty and is at most 1,024 characters.
Claude Code accepts more optional frontmatter fields, such as allowed-tools and disable-model-invocation. Those fields control invocation and tool access in Claude Code; other clients may ignore them.
Which skills are available, and who makes them?
The Claude help centre describes four types of skill on claude.ai, sorted by who wrote them. Claude Code adds a second axis: where the folder lives.
Type | Who makes it | Example |
|---|---|---|
Anthropic skills | Anthropic | Excel, PowerPoint, Word and PDF files, which Claude invokes automatically when relevant |
Custom skills | You or your team | Your brand voice, a report format, a review checklist |
Organisation skills | Team and Enterprise owners provision them for every user | A company style guide switched on by default |
Partner skills | Software vendors | Notion, Figma and Atlassian skills built to work with their MCP connectors |
Claude Code skills come from four places. Personal skills sit in ~/.claude/skills/ and apply across all your projects, project skills sit in the repository's .claude/skills/ folder and travel with the code, plugin skills arrive inside an installed plugin, and enterprise skills come from managed settings. When two share a name, enterprise wins over personal and personal over project.
Claude Code also merged custom slash commands into skills. A file at .claude/commands/deploy.md and a skill at .claude/skills/deploy/SKILL.md both create /deploy. Wherever a skill comes from, it works the same way once loaded.
How do skills compare with other Claude capabilities?
Skills provide procedure, and the other Claude capabilities provide context, connections or packaging. The help centre draws the same lines, and the table puts them side by side.
- Skill
Reusable instructions and supporting resources.
- MCP
A connection exposing supported tools.
- Task
The workflow can use both where appropriate.
Instructions and connected tool access solve different parts of the job.
Capability | What it gives Claude | When it is loaded |
|---|---|---|
Prompt | Guidance for one conversation | Only in that chat |
Custom instructions | Preferences for every conversation | Always |
Project | Background documents for one project | Always, inside that project |
Skill | A procedure for one kind of task | When a request matches its description |
MCP connector | Access to an outside system's data and actions | When Claude calls one of its tools |
Plugin | A package of skills, MCP servers and other parts | Once installed, each part behaves as above |
The pairing that matters most is skills with MCP. MCP connects Claude to your systems, and skills teach Claude what to do with them, which is how Anthropic framed it in its March 2026 explainer. MCP vs Skills works through when each one is the right tool.
From RankX AIReview the procedure before connecting more tools.Test it on one straightforward product and another with a missing material or contradictory size.Read the agent-skills referenceWhat is the Agent Skills open standard?
The Agent Skills open standard is the published specification at agentskills.io that defines the SKILL.md format, so one skill folder can work in many products. It was released in December 2025, and Claude Code's documentation says its skills follow it, with extras such as invocation control on top.
Adoption moved quickly. When we counted on 17 September 2026, the standard's client showcase listed 46 products, including OpenAI's ChatGPT and Codex, GitHub Copilot, Cursor and Gemini CLI.
Packaging followed. Agent Plugins 1.0.0, published by Vercel on 6 August 2026 with maintainers from AWS, Cursor, Microsoft, OpenAI and Vercel, bundles Agent Skills and MCP servers into one installable unit.
Where do skills run: claude.ai, Claude Code or the API?
Skills run on claude.ai and the Claude desktop app, in Claude Code, in the Claude Agent SDK and on the Claude API, but each surface keeps its own copy. Custom skills do not sync between claude.ai and the API: you upload custom skills through the Skills API for API use and through Customize on claude.ai.
Surface | How skills get there | Who can use them | Where scripts run |
|---|---|---|---|
claude.ai and desktop app | Upload a zip under Customize, then Skills | You, plus skills an organisation owner provisions | Anthropic's code execution sandbox |
Claude Code | A folder on disk, or a plugin | You, or everyone who clones the project | Your own machine, with its network access |
Claude API | Upload through the Skills API, then list the | Everyone in the API workspace | A sandbox with no network access |
On claude.ai, skills are available on the Free, Pro, Max, Team and Enterprise plans once code execution is turned on. The pre-built document skills work on claude.ai and the API but not in Claude Code.
Claude Code signed in with your Claude account picks up the skills from that account, and the help centre says this sync is one way: a folder you add in Claude Code never travels back. Signed in with an API key, Claude Code does not sync account skills, and the Agent SDK reads skills from the same folders as Claude Code.
How do you install and use a skill?
Installing a skill takes a minute on either surface. These steps cover claude.ai and Claude Code.
Turn on code execution
On claude.ai, open Settings, then Capabilities, and switch on code execution if it is off.
Open the skills list
Open Claude, then Customize, then Skills, to browse the directory or toggle the skills you already have.
Upload your own
Choose the plus button, then upload a zip that contains the skill folder with its SKILL.md inside.
Or drop a folder in Claude Code
Put the folder in .claude/skills/ for one project or in ~/.claude/skills/ for every project; Claude Code picks up changes without a restart.
Ask for the job
Describe the task in your own words and let the description match, or ask Claude to use the skill by name.
Skills without an outside system need no setup, but a skill that acts on one needs a working connection first. RankX AI's skills call the RankX AI MCP server, so connect it to Claude before you add them; the Agent Skills page has the complete list and the download for each.
For Claude Code, the docs give one command per skill, for example curl -o .claude/skills/rankxai-visibility-audit/SKILL.md https://rankxai.com/agent-skills/rankxai-visibility-audit.md after creating that folder.
How do you create skills of your own?
You create a skill by writing one Markdown file. Give it a name, a description that says when it applies, and instructions written the way you would brief a careful new colleague. The official skill authoring best practices add one rule worth knowing first: assume Claude is already smart, and write only what it could not guess.
Then test it on real requests before you trust it. How to write a Claude Skill that loads covers the description limits that differ by surface, the naming rules and a simple evaluation loop.
Which use cases and workflows suit a skill?
The best use case for a skill is any task you want Claude to do the same way every time. If you keep pasting the same brand guidelines, report format or checklist into chats, that text belongs in a skill.
Use case | What the skill holds |
|---|---|
Brand voice | Tone rules, banned phrases and approved product names |
Recurring reports | The metrics to pull, the order of sections and the chart to draw |
Content briefs | What a finished brief contains and how the target keyword is chosen |
Product page reviews | The fields to check and the questions to raise when a fact is missing |
Anthropic's launch post quoted Rakuten reporting finance workflows falling from a day to an hour with skills. That is one customer's account, not a benchmark, but it shows the pattern: the gain comes from a repeatable procedure.
For SEO work, skills become the method that turns a tool connection into a finished job. RankX AI publishes six skills for its MCP server: a visibility audit, keyword research, competitor analysis, site health, a content brief and shopping visibility, each with the confirmation steps already written in.
Running a visibility audit from inside Claude shows one in use. The MCP connection is part of the RankX AI web app for WooCommerce and other websites; the Shopify app works inside Shopify admin instead.
How do you check that a skill is safe?
A skill is as safe as its source, because it is instructions and sometimes executable code. The official guidance is to use skills only from sources you trust: ones you wrote yourself or got from a party you already trust.
Claude applies a skill's instructions as written, so a malicious skill can direct Claude to call tools or run code in ways that do not match its stated purpose. A skill that fetches content from outside URLs carries extra risk, because the fetched text can contain instructions of its own.
The audit is quick. Open every file in the folder, look for network calls or file access the task does not need, and remember that in Claude Code a script runs on your own machine.
Use a skill to keep product reviews consistent
A merchant can give Claude a skill with a repeatable product review: read the approved specifications, find missing buyer information, propose clearer copy and list every unresolved fact. The skill standardises that review even when the source is an authorised CSV export rather than a live store connection.
Test it on one simple product and on another with a missing material or a contradictory size. The useful result is a reviewable proposal with questions where the evidence is missing.
A skill does not grant permission to read a private catalogue or change a live product. A working connection, suitable access and an explicit decision to apply the proposal remain separate requirements.
Questions about AI Agents and MCP
Can I use Claude Skills for free?
Yes. The Claude help centre lists skills on the Free, Pro, Max, Team and Enterprise plans, with code execution turned on, and that includes uploading your own. Skills you download, like the six RankX AI publishes, are free files; whatever a skill operates on, such as a RankX AI account, is priced by that product, not by the skill.
How is a skill different from a prompt?
A prompt is guidance for one conversation, and you retype it each time. A skill is saved once, sits as a one-line description until a request matches it, then loads its full instructions on its own. Use a prompt for a one-off request and a skill for a task you repeat.
Can a skill work outside Claude?
Increasingly, yes. Anthropic published Agent Skills as an open standard at agentskills.io in December 2025, and when we counted on 17 September 2026 the standard's client showcase listed 46 products, including OpenAI's ChatGPT and Codex, GitHub Copilot, Cursor and Gemini CLI. A skill is a folder of Markdown and files, so the same skill travels between clients that read the format.
What is the difference between Claude Skills and plugins?
A skill is one unit of procedure: a folder with instructions for one kind of job. A plugin is a package; in Claude Code it can bundle several skills, MCP server connections, hooks and agents into one install, and the Agent Plugins standard of August 2026 made skills and MCP servers its portable core. Install a plugin and you may receive several skills at once.
Are skills safe to install?
Treat a skill like software, because it is instructions and sometimes code. Anthropic's guidance is to use skills only from sources you trust, since a malicious skill can direct Claude to call tools or run code in ways that do not match its stated purpose. Read every file in the folder before you turn it on.
Sources
- Anthropic, Introducing Agent Skills, 16 October 2025, updated 18 December 2025 (opens in a new tab) Checked 2026-10-29.
- Anthropic, Skills for organizations, partners and the ecosystem, 18 December 2025 (opens in a new tab) Checked 2026-10-29.
- Anthropic, Claude Skills explained, 5 March 2026 (opens in a new tab) Checked 2026-10-29.
- Claude Platform docs, Agent Skills overview: loading levels, limits, surfaces and security (opens in a new tab) Checked 2026-10-29.
- Claude Platform docs, Skill authoring best practices (opens in a new tab) Checked 2026-10-29.
- Agent Skills open standard specification, agentskills.io (opens in a new tab) Checked 2026-10-29.
- Agent Skills client showcase, agentskills.io: 46 products listed when counted on 17 September 2026 (opens in a new tab) Checked 2026-10-29.
- Claude Code docs, Extend Claude with skills (opens in a new tab) Checked 2026-10-29.
- Claude Agent SDK docs, Extend agents with skills (opens in a new tab) Checked 2026-10-29.
- Claude Help Center, What are Skills? (opens in a new tab) Checked 2026-10-29.
- Claude Help Center, Use Skills in Claude: plans, upload and one-way sync to Claude Code (opens in a new tab) Checked 2026-10-29.
- Claude Help Center, How to create custom skills (opens in a new tab) Checked 2026-10-29.
- Vercel, Introducing Agent Plugins 1.0.0, 6 August 2026 (opens in a new tab) Checked 2026-10-29.
CoversThis article covers the AI Agents and MCP topic, the Content Studio feature and the AI Readiness Score tool.
Terms usedContext Window, Token (LLM) and Training Data.
Read this page asMarkdown: /blog/what-are-claude-skills.md.
All articlesEverything RankX AI publishes is listed on the blog index.
Ask an assistantAsk ChatGPT (opens in a new tab), Ask Claude (opens in a new tab) or Ask Perplexity (opens in a new tab).
Preferred sourceIf Google is your front door, you can add RankX AI as a preferred source (opens in a new tab), which asks your own results to surface more of what we publish.


