A Claude Skill is a folder with a file called SKILL.md that contains a work instruction: when the task applies, which steps it runs through, and which rules have to be followed. Claude loads the Skill automatically as soon as a matching task comes up, or you call it with a slash command such as /seo-audit. Skills run in Claude Code and in the Claude app, on every plan. A Skill is not a subagent (a subagent works in its own context and only returns a summary) and not a Project (a Project holds knowledge, not processes). And you write Skills without programming: a Skill is plain text, exactly the way you would write a work instruction for a new employee.
I run my AI workforce in Claude Code. Every AI employee has its own folder with its own Skills, for example seo-audit and geo-audit for my SEO employee Sebastian, website-deploy for Ralf, lexoffice-rechnungen and neukunden-mail for Sonja. For me, Skills are one of the three reasons my entire workforce runs on Claude. The other two, and all my other guides, are on my Claude overview page.
What a Claude Skill is, and what it is not
Technically, a Skill is a folder with a file called SKILL.md. It describes, in plain text, when the Skill applies and how the task runs. Templates, reference material, and small helper scripts can be added. Claude only loads the Skill when it's needed and then works through the instruction step by step.
The crucial part isn't the tech, it's the format: a Skill is a readable document. You can open it, understand it, and change it like a work instruction for a person: no flowchart, no code you can't see through. Anthropic ships some Skills preinstalled, for example for Excel spreadsheets, Word documents, PowerPoint presentations, and PDFs. But the interesting Skills are the ones you write yourself.
Three things get mixed up with Skills all the time. The CLAUDE.md, or the Project, holds who your employee is, what it knows, and which rules always apply; it is loaded in every session. A Skill holds how one specific task runs, and it is only loaded for that task. A subagent is a separate workstation with its own context window. I once put it this way: a Skill is an employee for one task. The level above that is an employee with a workstation of its own. The precise distinction, with a decision rule, is in Claude Subagents, Skills, Agents; the file for permanent knowledge is covered in CLAUDE.md: Claude Code's Memory.
Installing Claude Skills: Claude Code, Claude.ai, marketplace
There are three ways to get Skills (as of September 2026, Skills are available on all plans including Free):
- As a folder in Claude Code. A personal Skill lives at
~/.claude/skills/<name>/SKILL.md and applies to all your projects. A project Skill lives inside the project at .claude/skills/<name>/SKILL.md and travels with the repository. Claude Code detects both automatically; you call them with /<name>.
- In the Claude app. Enable the code execution feature under Settings, then open the Skills section under "Customize." There you can switch on existing Skills or upload your own as a ZIP file. The folder contains the SKILL.md, and the folder name has to match the Skill name.
- Through a plugin marketplace. Skills can be bundled into plugins and distributed through a marketplace. You get Anthropic's own collection in Claude Code with
/plugin marketplace add anthropics/skills, then install individual packages with /plugin install. Third-party collections such as Superpowers are added the same way, using the GitHub path of the repository in question. Skills from plugins then show up with a prefix, in my setup for example as superpowers:brainstorming.
The path from installation to the first session is in Claude Code: A Guide for Beginners.
Creating Claude Skills: SKILL.md step by step
My advice: forget the tech for a moment and imagine you're writing a work instruction for a new employee. That's exactly what a Skill is. A good work instruction answers four questions:
- When does it apply? "Whenever a proposal draft is created."
- What are the steps? In order, concrete: first read the inquiry, then match the services, then write the draft, then check it against the checklist.
- Which rules apply? Phrases that always have to be included. Phrases that must never appear. What gets submitted for approval.
- What does a good result look like? One real example says more than ten rules.
The file itself starts with a short header of two fields, followed by the instruction as text. This is how the seo-audit Skill of my SEO employee begins (the description is in German, because the workforce works in German):
---
name: seo-audit
description: Use when Kevin eine Website bewerten lässt, komplettes SEO-Audit (technisch, On-Page, Local) mit Tool-Belegen und priorisiertem Maßnahmenplan. Ergebnis ist audit.md im Auftragsordner des Kunden.
---
The directory name is the name you call it by, so /seo-audit; the name field is optional according to the docs and only the display name in listings. The description field is the actual trigger: Claude uses it to decide whether the Skill fits the current task. Below the header, the Skill describes the process in prose: inventory, evaluation, action plan, review pass, and which tools are allowed for it. It contains no program code.
The easiest way to your first Skill: let Claude write it itself. Once you've walked through a task cleanly in a chat and the result is right, say: "Turn this process into a Skill so you work exactly the same way next time." Then proofread it, correct what's off, and call the Skill with /<name> next time.
Claude Skills examples: five from my workforce
Here are five real Skills from my workforce's repository:
If you don't have a workforce yet, these are the three Skills I would start with: a proposal Skill (summarize the inquiry, match services and prices, draft in your own voice, collect open questions, nothing leaves draft status without approval), a review Skill for anything you publish (check the draft against your own rules and flag violations instead of silently rewriting), and a weekly-report Skill (collect finished tasks, list open points, three suggestions for next week, always the same shape). The rule of thumb behind it: anything you could teach an intern in five minutes, you can write down as a Skill.
The pattern shows what makes a Skill usable: each one covers a recurring task, each one produces a nameable result, and each one ends at a point where I decide. That is a rule in my setup: a Skill prepares, I approve. The content folder my content editor-in-chief Conny works in holds more than 50 such Skill folders; she uses eight of them herself, among them /hook and /reel-skript. Every draft runs against my hook rules before I even see it. Which rulebooks sit behind those drafts is on Conny, who runs my Instagram channel.
Claude Skills on GitHub: Superpowers and collections
Anthropic maintains the anthropics/skills repository with example Skills, the specification for the Skill format, and a template. Superpowers by Jesse Vincent is a third-party Skill package with a complete working methodology for coding agents: brainstorming, systematic-debugging, test-driven-development, writing-skills, and more. I use some of them in my workforce alongside our own Skills.
To be honest about it: a third-party Skill is a work instruction that Claude carries out on your behalf, possibly with access to your files. Read every SKILL.md before you install it, and what was written for a development team doesn't automatically fit a consulting practice.
Sharing Skills with your team
The simplest way is Git: a Skill in the project folder under .claude/skills/ gets committed with the repository. That is how it works for me: the Skills of all my AI employees live in one shared repository. Why those files also survive a switch of tools is in CLAUDE.md vs AGENTS.md: What Survives. The second way is a plugin of your own with a skills/ directory that you distribute through a marketplace. In the Claude app, Team and Enterprise organizations can provision Skills for all users or share them with specific colleagues.
Three mistakes that make Skills useless
After many work instructions written and thrown away again, I know the typical traps:
- Too vague. "Write good proposals" isn't a Skill, it's a wish. A Skill needs verifiable steps and rules. If you can't tell whether Claude followed the instruction, it's worded too softly.
- Too long. A Skill that tries to cover every edge case becomes a novel nobody maintains anymore. Write down the normal case and only add edge cases once they actually occur.
- Skills for one-off tasks. A Skill only pays off for processes that recur. For a one-time task, the chat is enough.
Frequently asked questions
Are there Claude Skills in languages other than English?
Yes, a Skill is plain text in whatever language you write it in. My Skills are written in German; only the description header line starts with the English "Use when" in some of them, because that is what Claude recognizes as the trigger.
Do Skills work in Claude.ai or only in Claude Code?
Both. In Claude Code they live as folders; in the Claude app you upload them as a ZIP after enabling code execution. I tried the ZIP route myself in Claude Cowork: built the Skills, uploaded them, and it worked well. Day to day I have still worked almost exclusively in Claude Code for the past three months or so, because that is where my employees' folders are.
Do Skills cost extra?
No, Skills are part of every plan, including the free one (as of September 2026, according to Anthropic's help article). What a loaded Skill does cost is space in the context window: its text counts toward your plan's usage. So keep it short. The plans themselves are covered in Claude Pricing and Plans Explained.
Skill or Project instruction: which one do I use when?
Everything that should always apply (who the employee is, tone, prohibitions) belongs in the role file or CLAUDE.md; when starting inside the Claude app, it belongs in the Project instructions instead. Everything that only applies to one specific task (steps, checklist, output format) belongs in a Skill. How the role principle works in Claude Projects is in How to Use Claude Projects Right.
How many Skills make sense?
As many as you have recurring tasks, and not one more. For most of my employees it's a handful: Sebastian has two, Susi seven. The exception is the content area with more than 50 Skill folders, because it serves so many formats. A Skill you haven't called in months can go.
Where does the SKILL.md live?
In Claude Code, either at ~/.claude/skills/<name>/SKILL.md for you personally or at .claude/skills/<name>/SKILL.md inside the project. In the Claude app, it sits inside the uploaded ZIP file.
How do I create a Claude Skill?
With a single file. Anthropic's Skills documentation (retrieved September 21, 2026) puts it this way: "Every skill needs a SKILL.md file with two parts: YAML frontmatter between --- markers that tells Claude when to use the skill, and markdown content with the instructions Claude follows when the skill runs." The folder name becomes the command and the description becomes the trigger, verbatim: "The directory name becomes the command you type, and the description helps Claude decide when to load the skill automatically." So a Skill needs neither code nor a build tool: create the folder, write the SKILL.md, done. The four questions the instruction inside should answer are in the section on creating Skills above.
What are some Claude Skills examples?
The five in the table above are real Skills from my workforce: seo-audit and geo-audit for Sebastian, website-deploy for Ralf, lexoffice-rechnungen and neukunden-mail for Sonja, plus /hook and /reel-skript from Conny's content folder. What they share is the pattern: a recurring task, a nameable result, and a decision that stays with me. If you want an example to copy from, take the template in the anthropics/skills repository, and for your first own Skill the three suggestions above: proposal, review before publishing, weekly report.
How do I install a Claude Skill, and where does it apply?
In Claude Code you install a Skill by putting its folder in one of two places. According to the docs (retrieved September 21, 2026), a personal Skill at ~/.claude/skills/<skill-name>/SKILL.md loads in "All your projects on this machine, but not Cowork or cloud sessions". A project Skill at .claude/skills/<skill-name>/SKILL.md loads in "Sessions in this repository. Commit it so your team gets it too". For the Claude app and for plugins, the other two routes are in the installation section above.
Where to start
Take a task you're doing with AI this week anyway and walk through it once cleanly in a chat until the result is right. Then have Claude turn it into your first Skill, proofread it, and put it under .claude/skills/. That's all it takes to get started. Why exactly this combination of Skills, files, and memory brought me to Claude is covered in Claude Over ChatGPT: My AI Employees.
If you want to set the whole thing up properly from the start, with a role, personnel file, and onboarding instead of individual Skills: the complete process is in Hiring an AI Employee: The Process. The order in which Skills, Projects, and Claude Code build on each other is in Learn Claude: The Path in Five Stages. The eight packages are ready to use in my community.
Sources: Skills in Claude Code (Anthropic docs), Using Skills on Claude (Anthropic help), anthropics/skills on GitHub, Superpowers on GitHub, all retrieved in September 2026.
