What Are Claude Skills? A Marketer's Guide

What are Claude Skills?

A Claude Skill is a folder containing a SKILL.md file: YAML frontmatter carrying a name and a description, then a markdown body of instructions Claude follows, optionally alongside scripts, reference material and templates. Anthropic's engineering team puts it plainly, skills are folders of instructions, scripts and resources that agents discover and load dynamically, and offers the analogy that building one is like writing an onboarding guide for a new hire. That is the accurate mental model: a skill does not give Claude new powers, it is a file that tells Claude exactly how you do a specific task.

Anthropic announced Agent Skills on 16 October 2025 and published the format as an open standard that December. Skills are reusable by design: write the procedure once and Claude applies it in every conversation where it fits, which is the difference between a skill and the prompt you keep retyping.

How do Claude Skills work?

The economics rest on progressive disclosure, in three levels. At the start of a session Claude holds only each skill's metadata, the name and description from the frontmatter, roughly a hundred tokens riding along in the system prompt. When your request matches a description, Claude reads the SKILL.md body, and only if the job demands them does it open the bundled files, so Claude loads what it needs and nothing else. Expertise sits on the shelf costing almost no context window until the moment it applies.

When a skill bundles code, Claude runs it through the code execution tool in a sandboxed container on claude.ai and the API, or directly on your machine in Claude Code. That is how the document skills produce real files rather than descriptions of files: the instructions say what to build, the executable scripts build it.

How does Claude decide which skills to use?

By matching your request against the descriptions of every available skill, automatically when relevant: ask for a quarterly report as a spreadsheet and the Excel skill loads without being named. You can also direct Claude explicitly, use the visibility audit skill, which is the reliable route when several skills overlap on the same use case. Claude decides whether to load; the description decides how well it can, which is why the description is the one line worth writing carefully in a skill of your own.

What types of Claude Skills exist?

Three types, by who wrote them and where they live. First-party skills ship with the product: the document skills that produce real Excel, PowerPoint, Word and PDF files are on by default. Custom skills are yours: uploaded to claude.ai as a zip, or provisioned programmatically, since developers can upload custom skills through the Skills API to use with the Claude API's code execution tool. And vendor skills come with products you use, the way RankX AI ships six for SEO work.

Claude Code skills add one more axis: location. Personal skills live in your home skills directory and apply across all your projects; project skills live in the repository's own skills folder and travel with the code; plugin skills arrive bundled inside installed plugins. All of them work the same way once loaded; the location only decides who else gets them.

What can a marketer actually do with skills?

The pattern that matters is turning your standards into defaults. Anything you find yourself re-explaining to an AI assistant, brand guidelines and voice, report formats, your team's best practices for qualifying a keyword, what a finished brief contains, belongs in a skill, written once and applied every time. The launch customers point the same way: Rakuten reported finance workflows falling from a day to an hour.

For SEO work specifically, skills are how a tool connection becomes a job. RankX AI ships six alongside its MCP server: a visibility audit, keyword research, competitor analysis, site health, a content brief workflow and shopping visibility, each a written procedure with the confirmation steps and honest denominators already in it. Running the visibility audit end to end shows what one looks like in use, and the Agent Skills reference documents all six.

How do you create your own Claude skill?

Write a markdown file. A minimal skill is one SKILL.md: frontmatter with a name and a description that says when the skill applies, then the instructions Claude follows, written the way you would brief a careful new colleague, because a skill teaches Claude by explanation rather than code. Keep the description specific, keep the body concise, and add scripts or reference material only when the job needs them. Then test it the obvious way: ask Claude for the job and check whether the skill loads and whether the result matches what you meant. If you want Claude to run the same judgement your team applies, write the judgement down; that is the entire trick.

Where do you find and install skills?

In claude.ai, open Settings, then Customize, then Skills: browsable skills toggle on there, and custom skills upload as a zip. The built-in document skills are on by default, and Team and Enterprise admins can provision skills centrally. In Claude Code, a skill is a directory: put the folder at .claude/skills/ in a project or ~/.claude/skills/ for every project, and invoke it by name or let Claude match it. RankX AI's skills install with one download each, for example curl -O https://rankxai.com/agent-skills/rankxai-visibility-audit.md placed into a folder of the same name.

One honest caveat about scope: a skill needs a working connection to act on anything external. The RankX AI skills call the product's MCP tools, so connect the server first, then add the skill files. Skill stores also do not sync across surfaces: a skill uploaded to claude.ai and a folder in Claude Code are separate installations of the same file.

Where do skills run, and what keeps them safe?

Skills work across claude.ai, the Claude desktop app, Claude Code, the Claude Agent SDK and the API, where bundled code runs inside a sandboxed code-execution container. In Claude Code, scripts run on your own machine, which is power and responsibility in equal measure. The official security guidance is blunt: skills give Claude capabilities through instructions and code, a malicious one can steer tool use against you, so install only from trusted sources and read what you install. A SKILL.md is plain markdown; the audit takes minutes.

How do skills relate to MCP?

They are complements with a clean division: MCP connects an assistant to your systems, and skills teach it what to do with them, the vendor's own framing since the March 2026 explainer. The pairing is now formalised in the ecosystem too: Agent Plugins, a packaging standard launched in August 2026 by Amazon, Cursor, Microsoft, OpenAI and Vercel, bundles skills and MCP servers into one installable unit. MCP vs Skills works through when each is the right tool, and how MCP works covers the connectivity half.

Sources

• Anthropic, Introducing Agent Skills, 16 October 2025, checked 20 Aug 2026
• Anthropic engineering, Equipping agents for the real world with Agent Skills, checked 20 Aug 2026
• Agent Skills open standard specification, checked 20 Aug 2026
• Anthropic, Skills for organizations, partners and the ecosystem, 18 December 2025, checked 20 Aug 2026
• Anthropic support centre, What are Skills, checked 20 Aug 2026
• Vercel, Introducing Agent Plugins 1.0.0, 6 August 2026, checked 20 Aug 2026