Claude Skills explained: what they are, how to install and share them (2026)

If you give Claude the same briefing every week, you are losing time a skill could save. A Claude Skill captures a way of working once, and Claude brings it up on its own when a task calls for it. A lot has changed since the launch in October 2025: you now install skills through Customize > Skills, on Team and Enterprise you share them with colleagues, and plugins bundle them with connectors into a single package.

In this article we explain what skills are, how to install, share and build them, and what we learned from using them in practice. Everything was checked on 3 October 2026. If you want the bigger picture of the app first, read our guide to Claude Desktop.

What are Claude Skills and why are they more than a system prompt?

Claude Skills are folders of instructions, scripts and resources that Claude loads dynamically to perform better at a specialised task. A system prompt or project instruction is always active. A skill only comes into play when the task calls for it, and it works everywhere in Claude. That way you capture a way of working once and never have to explain it again in every conversation.

Think of a consultant walking in for the first time and an employee who has been around for three years. Equally smart. The employee is more productive, because the context is already there.

The difference from the other Claude features has practical consequences (Claude Help Center):

Project instructions are the onboarding document for a new employee. A skill is the SOP for one task, the answer to “how exactly do I do this?”. For a broader comparison of the AI toolset, see our practical guide to ChatGPT, Claude and Perplexity.

Anthropic launched Skills on 16 October 2025 for the Claude apps, Claude Code and the API (Anthropic). On 18 December 2025 the specification became an open standard, Agent Skills, with the same reasoning as the Model Context Protocol: a skill should be able to move between tools and platforms. In the coverage of that step, Lian Jye Su, Chief Analyst at Omdia, told AI Business: “The agent can just resolve the task on its own, and in that case, the agent becomes a lot smarter, a lot easier to deploy.”

Skills work today on Free, Pro, Max, Team and Enterprise. In the API they have been out of beta since 19 August 2026. Skills you have enabled in Claude also work in the add-ins for Excel, PowerPoint, Word and Outlook.

ClickForest helps Flemish SMEs in Mechelen and Antwerp with AI for growth, and a skill is a good starting point for that: capturing one way of working that today still lives in someone’s head.

How does a SKILL.md file work technically?

A skill is a folder with at least one file: SKILL.md. At the top sits YAML frontmatter with a name and a description, below it the instructions in markdown. Claude first reads only that name and description. It loads the full instructions when the skill is relevant, and it only opens scripts or reference files when the task really needs them.

Anthropic describes this as three levels (Agent Skills documentation). Name and description are always in context, about 100 tokens per skill. If a skill looks relevant, Claude loads the instructions from SKILL.md, which should ideally stay under 5,000 tokens. Scripts and reference files only follow when the task asks for them. When Claude runs a script, the code itself never enters the context window: only the output counts.

A minimal example:

---
name: seo-product-description
description: Writes SEO-optimised product descriptions with search intent, primary keyword in H1 and first paragraph. Use for Shopify product pages or e-commerce content.
---

# SEO product description
Always write with BLUF structure...

The description is the most important part of the whole skill, because Claude uses it to decide whether the skill ever gets used. Too vague, and it never loads. Too broad, and it shows up at the wrong moment. In claude.ai the description can be 200 characters long, in the open standard and the API 1,024 characters. Claude Code truncates the description and the when_to_use field together at 1,536 characters in the skill listing.

That simplicity is what got Simon Willison excited on launch day:

“Skills are Markdown with a tiny bit of YAML metadata and some optional scripts in whatever you can make executable in the environment. They feel a lot closer to the spirit of LLMs—throw in some text and let the model figure it out.”

— Simon Willison, independent open source developer and creator of Datasette (simonwillison.net, 16 October 2025)

In Claude Code a skill can live in five places (Claude Code documentation):

• Personal: ~/.claude/skills/, available in all your projects
• Project: .claude/skills/ in the project folder, included in version control
• Enterprise: centrally managed through managed settings
• Plugin: skills that come with a plugin, invoked as /plugin-name:skill-name
• Synced: the skills from your claude.ai account, in ~/.claude/skills/synced/

When two skills share a name, Enterprise beats personal, and personal beats project. Old custom commands keep working: Anthropic merged them into skills.

When is a Claude Skill the right choice?

A Claude Skill is the right choice when you perform the same task repeatedly and expect consistent output every time that meets fixed criteria. The practical test: if you repeat the same briefing in every conversation, you have a skill in the making. For one-off questions and free-form creative work, a sharp prompt is enough.

Situations where skills work well:

• Weekly report: Claude knows which sections belong in it (summary, KPIs, action points) and which formatting you use.
• SEO audit: a fixed checklist for title tags, meta descriptions, H1, internal links and schema, useful alongside an ongoing SEO programme.
• Product descriptions: tone of voice, structure and keywords are fixed for every new Shopify upload.
• Ad copy: headlines and descriptions that stay within the character limits.
• AI chatbot: the knowledge boundary, response style and escalation rules of your AI chatbot stay the same in every interaction.
• Web design: the official frontend-design skill gives Claude a design direction for projects around building with AI.

For ad-hoc questions and one-off assignments, good prompts are enough. In that case, start with how to write better prompts.

Arun Chandrasekaran of Gartner summed up the promise when the open standard was announced:

“For many of the business use cases that require human input, now you can actually have an AI model with the skill go ahead and do that action on your behalf.”

— Arun Chandrasekaran, analyst at Gartner (AI Business, 18 December 2025)

Which Claude Skills should you definitely install?

Start with Anthropic’s own skills. The document skills for Word, Excel, PowerPoint and PDF already run automatically in Claude, and skill-creator helps you build your own skills. The official GitHub repository holds 19 skills (as of October 2026), intended as demonstrations and learning material. So test them in your own environment before you let critical work depend on them.

The most useful official skills for marketing, content and web work (anthropics/skills on GitHub):

The README is clear about it: the skills are there “for demonstration and educational purposes only”. Only the four document skills actually run in production. Outside Anthropic the range is growing through plugins, such as Salesforce in Claude with 37 sales skills (beta since 15 September 2026, release notes).

Where do you find community skills and how do you install them?

The most reliable routes today are the skills directory in Customize > Skills, the Claude Marketplace and the plugin marketplaces in Claude Code. Example skills are on GitHub in anthropics/skills, and agentskills.io shows which tools support the open standard. In claude.ai you install skills through Customize > Skills; in Claude Code you put the skill folder in a fixed folder on your computer.

The directory in the app. Go to Customize > Skills, click the ”+” button and choose “Browse skills”. There you find skills from partners such as Notion, Figma and Atlassian (Claude Help Center). Skills you install from the directory can be used but not edited.

“Now Claude doesn’t just see Jira tickets or Confluence pages, it knows what to do: turning specs into backlogs, generating status reports, surfacing company knowledge, triaging issues, and more.”

— Josh Devenny, Head of Product, Rovo Skills at Atlassian (Claude blog, 18 December 2025)

The Claude Marketplace. Since 23 September 2026 Anthropic has brought connectors and plugins together on a dedicated website, with more than 2,000 items (as of October 2026), including ones from Atlassian, Google, Microsoft, Notion and Salesforce (Anthropic).

Plugins. A plugin bundles skills, connectors and sub-agents into one package and works on all paid plans. Since 25 September 2026 Anthropic has called plugins “the main way to build third-party extensions for Claude” (Anthropic). Standalone skills and connectors remain building blocks, and the directory keeps listing them. Note that hooks and sub-agents from a plugin run in Cowork and Claude Code, not in regular chat (Claude Help Center).

Plugin marketplaces in Claude Code. In your first interactive session, Claude Code adds Anthropic’s official plugin marketplace, which is separate from the Claude Marketplace on claude.com (Claude Code documentation).

GitHub and agentskills.io. Besides the anthropics/skills repository, agentskills.io lists 46 tools that support the open standard (as of October 2026), including Cursor, GitHub Copilot, VS Code, Gemini CLI and Claude Code. A skill you build for Claude should in principle work there too.

Installing in claude.ai and the desktop app. First turn on “Code execution and file creation” under Settings > Capabilities. Then go to Customize > Skills, click the ”+” button, then “Create skill”, and choose “Upload a skill” to add your own skill as a zip file. On Team and Enterprise, an owner enables skills in Organization settings > Plugins & skills (Claude Help Center).

Installing in Claude Code. Put the skill folder in ~/.claude/skills/ for all your projects or in .claude/skills/ for one project. Claude Code watches those folders and picks up a new or changed skill in the running session, without a restart. Since mid-September 2026 Claude Code also syncs the skills and plugins from your claude.ai account: when a session starts and then roughly every 10 minutes. That sync is one-way. If a local skill has the same name as a synced one, the local one wins.

It doesn’t work the other way round. Cowork and cloud sessions don’t read your local ~/.claude/skills/, so a skill that only lives on your computer has to go into Customize > Skills before Cowork can use it. How Cowork has been merging with chat since 16 September 2026 is covered in Claude Cowork explained.

From skill to plugin. Add a .claude-plugin/plugin.json to a skill folder and Claude Code loads that folder as a plugin.

Which tasks in your marketing or e-commerce are ready for a skill? ClickForest helps Flemish SMEs and scale-ups in Mechelen and Antwerp with AI for growth, from the first skill to a working workflow. Get in touch.

How do you share Claude Skills with your team?

On Team and Enterprise you share a skill from within Claude itself: with specific colleagues, on Enterprise also with groups, or by publishing it to the organisation library. An owner can also give skills to everyone centrally. On Free, Pro and Max that isn’t possible. There, everyone installs the skill in their own account, or you share the folder through Git in Claude Code.

There are four ways to do it (Claude Help Center):

Sharing is on by default on Team and on Enterprise organisations that haven’t chosen a setting; in regulated configurations such as HIPAA it is off. Publishing is open by default on Team and starts off on Enterprise. On 2 October 2026 that changed for Enterprise organisations without a setting of their own: publishing switches to “Requires review”, and the security scan for skills and plugins turns on.

An owner can also switch off user-created skills entirely. Every sharing action shows up in the Enterprise audit log and the Compliance API, though the content of the shared skill does not.

How do you build your own Claude Skill?

The fastest route is skill-creator: you describe what you want, Claude asks questions and builds the folder, the SKILL.md file and the accompanying files. You can also write a skill by hand, or on a Mac demonstrate a way of working in Cowork with Record a skill. Afterwards, always test with three to five questions, including questions the skill should not respond to.

Method 1: skill-creator (no technical knowledge required)

1. Turn on the example skills in Customize > Skills
2. Type: “Use skill-creator to create a skill for [your task]”
3. Answer the questions skill-creator asks about your task
4. Claude generates the folder structure, the SKILL.md file and the bundle

skill-creator also improves existing skills and runs test cases to measure how well a skill performs (anthropics/skills).

Method 2: writing it by hand

1. Create a folder, for example ~/.claude/skills/your-skill-name/
2. Write SKILL.md with YAML frontmatter and instructions, with a description of at most 200 characters if you also want to upload the skill to claude.ai
3. Test with three to five representative questions and compare the output with and without the skill
4. Refine the description if the skill triggers at the wrong moment

Method 3: Record a skill

On Pro, Max and Team you can record a way of working in Cowork in the Mac app while you demonstrate it, for up to about 10 minutes. Claude turns that recording into a skill proposal. On Windows, in chat and on Free and Enterprise, this option isn’t available.

Tips that determine quality:

• Also say what Claude must not do (“Don’t sugarcoat the findings”, “Don’t skip the financial analysis”)
• Add examples of good input and output, as Anthropic recommends
• Keep SKILL.md under 500 lines and move detailed material to separate files (Anthropic’s best practices)
• Build several small skills rather than one big one: according to Anthropic, focused skills compose better

According to Anthropic’s enterprise guide, those test questions cover three kinds of cases: the skill should trigger, it should stay silent, or it’s a borderline case. Since early September 2026, /skill-doctor in Claude Code shows what each skill costs in context and how often it gets used.

How do you chain multiple skills together?

Claude combines multiple skills on its own when a task calls for it. Anthropic notes, though, that a skill can’t explicitly call another skill. If you want a fixed order, where the output of one step becomes the input of the next, you let a command or an agent orchestrate the steps. Each skill then stays small and sharp.

That is how it works at ClickForest for blog articles, with separate AI agents for the research and the checks. One command starts the keyword research and the fact research at the same time. Writing comes next, and before anything is shown, a separate agent checks every figure and every quote against the source. How that kind of workflow fits into an SEO approach is covered in combining SEO and AI.

Anthropic saw that idea from the start:

“Instead of building fragmented, custom-designed agents for each use case, anyone can now specialize their agents with composable capabilities by capturing and sharing their procedural knowledge.”

— Anthropic, engineering blog on Agent Skills (anthropic.com, 16 October 2025)

The broader shift towards agents that skills are part of is described in our article on AI agents and marketing.

What does working with Claude Skills teach you in practice?

ClickForest uses skills for recurring work: web copy, blog articles, LinkedIn posts, emails and structured data. What we learned along the way: keep your skills in one place, under version control. Don’t turn every skill on everywhere, because every description costs context. And let one skill with the facts about your business be the base that the other skills read from.

One source, under version control. If you edit a skill in several places, the copies drift apart sooner or later. That is why we edit skills in one place under version control and then replace the whole version in claude.ai. The “Edit with Claude” button in claude.ai is handy for a quick change, but it quietly turns claude.ai into a second source.

Every skill costs context, even when you don’t use it. The name and description of every active skill sit permanently in the context, about 100 tokens per skill according to Anthropic. With fifty skills that is already around 5,000 tokens before you have asked anything. So only turn a skill on where you need it. In Claude Code, /skill-doctor shows which skills you rarely use.

One skill with the facts about your business. Services, positioning, tone and the most important pages of your site: put them in one skill and have the other skills read from it first. When something changes, you update it in one place, and a skill that writes copy no longer has to guess.

ClickForest helps Flemish SMEs in Mechelen and Antwerp build a set of skills like this themselves through AI training, and supports them with marketing automation where skills carry the work.

How do you use Claude Skills safely?

Treat a skill like software: read it before you turn it on, even when a colleague shares it. A skill can run code and steer Claude, and according to Anthropic prompt injection and data exfiltration are the biggest risks. So only install skills from trusted sources, and never roll out a skill from an unknown source without a full review.

In its enterprise guide to Agent Skills, Anthropic gives a list of risk indicators to weigh every skill against before you approve it:

Since 6 August 2026 Enterprise organisations have had a scan (launched in beta) that checks third-party skills and plugins for malicious content when someone uploads or edits them. Since 2 October 2026 that scan has been on by default for Enterprise organisations that hadn’t set it themselves. The result is pass, warn or fail, and according to Anthropic a pass is no guarantee that a skill is safe in every respect. Skills added before the scan was turned on are not scanned, nor are MCP servers and hooks, and the scan doesn’t cover the API (Claude Help Center).

Be careful with plugins too: they can contain local MCP servers that run on your computer with the same permissions as any other program. If you use skills through the API, log which skills were in a request yourself, because the Skills API offers no usage statistics.

The same goes for Claude Cowork and AI agents: minimal access, and test first on work where nothing can break.

Conclusion: start small, build deliberately

Claude Skills turn a powerful general-purpose model into an assistant that knows your processes. Since the launch, Claude has gained sharing on Team and Enterprise, plugins, a marketplace and a sync between claude.ai and Claude Code.

Even so, it still starts with one skill for one recurring task. Build it well, test it and only then expand to chains. If you build too many skills too fast without sharpening the underlying way of working, you get a more complicated version of the same problem. And read every skill from someone else before you turn it on.

ClickForest helps Flemish SMEs and scale-ups in Mechelen and Antwerp put Claude Skills to practical use, as part of a broader approach with GEO, building with AI and AI tools by task. Want to know what the first meaningful step is in your situation? Book a no-obligation discovery call.