Claude skills are reusable instruction sets you give Claude for specific tasks, such as formatting a document in your house style or building a report according to a fixed method. Instead of repeating those instructions in every prompt, you capture them once in a skill file that Claude loads automatically when it is relevant.
The feature is most useful for recurring work where consistency and quality matter more than creativity. A skill does not turn Claude into a better generalist. It turns Claude into a specialist in exactly the type of work you do often.
Claude Skills is the name Anthropic uses for this feature. In practice you will hear both “a Claude skill” (singular, one instruction set) and “Claude skills” (plural, the feature as a whole).
What is a Claude skill?
A Claude skill is a folder containing a SKILL.md file that gives Claude instructions for a specific task. The file holds a short metadata description (a name plus a description of when the skill should be used), followed by the actual instructions in markdown.
Additional files such as scripts, templates or reference documents can sit in the same folder. They are only loaded when Claude actually needs them.
A skill is therefore essentially a documented way of working, and that way of working has to exist before it is worth capturing.
Skills are an open standard that Anthropic published at agentskills.io in December 2025 and handed over to the Agentic AI Foundation on 22 October 2026. In principle, the same skill therefore works on any AI platform that supports the standard, not just on Claude.
In practice this makes skills more future-proof than a purely Claude-specific feature. The effort you put into your own skill today stays useful if you later switch to another tool or use several side by side.
You can upload your own skills in claude.ai on every plan, including Free, as long as code execution is switched on. The built-in document skills run in the background as soon as you ask for something they fit. Skills that run scripts count towards your usage limit, so for heavy use it helps to have enough headroom in your plan.
How does a Claude skill get activated?
The way a skill is activated differs markedly from a normal prompt or a project instruction. In every conversation, Claude loads only the name and short description of each installed skill into the system prompt. The full content of the skill stays out of working memory until Claude determines that the skill is relevant to the request.
This principle is called “progressive disclosure”, and it is exactly what makes skills scalable. You can have a dozen skills installed without filling up working memory, because only the active skill is loaded in full at the moment it is needed.
Only the active skill takes up working memory, so you can install many skills side by side.
A skill for PDF forms is only read in detail when Claude has to fill in a PDF form, not for every other question you ask.
Activation works through semantic matching. Claude compares your request with the skill descriptions and selects which skill (or skills) apply. A request such as “make an Excel overview of these figures” matches the Excel skill; a request such as “write a Word document with this content” activates the Word skill. If no skill matches, Claude works as usual without skill context.
Skills are also composable. For a task that spans several steps, Claude can use more than one skill at once: it picks the right skill for each step and combines them within the same workflow. A skill’s instructions can also point to a follow-up step that another skill is suited to. There is no fixed coupling in which one skill calls another as a function; the combination arises because Claude activates the best-fitting skill for each step.
Which built-in skills does Claude have?
Claude has a number of built-in skills that Anthropic ships as standard and that are active as soon as code execution is switched on. These skills run in the background and need no separate installation.
The pre-built skills cover Word (docx), Excel (xlsx), PowerPoint (pptx) and PDF. The Excel skill builds complete spreadsheets with formulas, pivot tables and charts. The Word skill produces formatted documents with proper structure, headings and tables.
The PowerPoint skill creates presentations with slides, layouts and visuals. The PDF skill can fill in forms, merge files and extract data from existing PDFs. In practice you hardly notice these skills as a separate feature: they switch on automatically as soon as you ask for something they are suited to.
You do notice the effect in the output. An Excel file that Claude builds with the xlsx skill works in Excel itself, rather than being a text table you still have to format. A PDF form filled in via the pdf skill has the right fields in the right places; without the skill you would get new text that you then have to put into the PDF yourself.
When should you create your own skill?
Creating your own Claude skill makes sense once you have a specific way of working or company standard that Claude should apply consistently. Three scenarios come up again and again in organisations that actively use skills.
The first scenario is capturing a house style. A skill that describes your organisation’s house style (colours, fonts, formatting rules, standard text blocks) makes every document Claude produces look the same, without you having to explain it again in every prompt.
This goes further than a project instruction, because the skill also works outside the project in which it was created.
The second scenario is encoding a way of working. Sprint planning, code review, client onboarding, contract analysis, recruitment procedures: these are all tasks that follow a fixed structure and can be captured in a skill.
When Claude receives a request that fits within that structure, the skill activates and Claude works through the steps in the way you prescribed.
The third scenario is domain-specific knowledge. A skill can contain reference material such as a product catalogue, a list of regulations that apply to your sector, or a set of legal templates.
Claude consults that material only when the skill is activated, which improves accuracy and keeps context use low at the same time.
How do you create your own Claude skill?
Creating a Claude skill starts with writing a SKILL.md file. This is a markdown file with a small block of YAML metadata at the top (enclosed by three dashes above and below, containing, among other fields, a name and a description), followed by the instructions in plain markdown.
The description line is where most of the gain lies. It determines whether Claude activates the skill at all, because matching is based on that text. A vague description such as “helps with documents” leads to wrong activations or missed ones.
A strong description names three elements: what the skill does, when it should be used, and in which cases it should not.
Anthropic’s PDF skill, for example, explicitly names filling in forms and bulk processing as cases in which the skill should activate, and simple PDF viewing or basic conversions as cases in which it should not, to prevent false triggers.
The instructions below it describe what Claude should do once the skill is activated. Keep them structured: a short introduction, then steps or rules, and optionally a section with examples and a section with mistakes to avoid.
As with good project instructions, short and specific works better than long and rambling. A role statement at the top of the instructions strengthens the effect.
The same principles as for assigning a role in a prompt apply here: the more sharply you define the role and the task, the more consistent the output.
In claude.ai you upload a skill via Customize > Skills, where you choose + and Create skill, select the Upload a skill option and add the skill as a zip file. Switch on code execution beforehand under Settings > Capabilities. In Claude Code you place the skill folder directly in the project structure on the file system, by default under .claude/skills.
Because the format is the same everywhere, you build a skill once and use it wherever you need it: the same SKILL.md you develop in Claude Code can then be added to claude.ai by uploading the folder as a zip.
Anthropic also offers a built-in skill-creator that helps you generate a skill through a conversation, with an automatic review of the description and the structure. If you are making your first skill, that saves a lot of trial and error.
For Team and Enterprise plans there is also a marketplace that offers skills, plugins and connectors bundled together. An administrator can then roll out a complete work package for a role or department in one go, instead of installing separate skills for each employee.
You write a skill in the language you work in. Claude handles languages other than English without a separate setting, so a skill written in, say, Dutch or German activates just as reliably as one written in English.
What is the difference between Skills, Projects, Cowork and Artifacts?
Skills, Projects, Cowork and Artifacts are often mentioned in the same breath, but they solve four different problems. A quick comparison helps you decide which feature suits what you want to achieve.
A Claude project is a static workspace in which all chats share the same instructions and knowledge base. The context is always active within the project and always fully loaded. A skill works very differently: it is dynamic and only activates when a specific task comes up, including outside projects.
Claude Cowork is not an instruction feature but an agent: it carries out tasks and delivers files as output. Since 6 October 2026 it runs in the cloud by default on Pro, Max and Ultra, and it can use skills as a tool along the way. Since 16 September 2026, Cowork has been merged step by step with the regular chat for Pro and Max, so those agent functions become available from any conversation. A Claude artifact is not a workspace, instruction set or agent but an output format: a block of working output that appears in its own window in claude.ai and that you can share or remix, wherever and with whatever you made it.
The practical dividing line lies in scope and activation. Use a Project when you want to run a series of chats on the same topic with shared context. Use a Skill when you want to capture a way of working that applies across several projects and topics.
Use Cowork when Claude actually has to carry out the work, with tangible files as the result. Choose Artifacts when Claude delivers something within a chat that appears in its own window and that you want to share with others as a working result. The four are not mutually exclusive: you can happily combine projects, skills, Cowork and artifacts.
A skill you create also works outside the chat, for example when you use Claude Tag in Slack and a colleague calls up the same way of working.
Limitations of Claude skills
The limitations of Claude skills come down to two practical matters: the separation between environments and the maintenance of skills you share with others. Sharing a skill within an organisation is now possible.
Synchronisation is the most tangible issue. A skill you upload to claude.ai is not automatically available in the API, and a skill you upload via the API does not appear in claude.ai. The gap to Claude Code is smaller: from version 2.1.273, Claude Code loads the skills you have enabled in claude.ai, including skills an owner makes available to the organisation. That synchronisation only works one way, and not when Claude Code is signed in with an API key or runs through a cloud provider such as Amazon Bedrock.
If you also use skills in the API or through such a cloud provider, you have to upload the skill there separately or handle it through your own distribution system. It is not a blocker, but it is a maintenance burden that is easy to underestimate.
The second point is the maintenance of shared skills. Skills contain rules and knowledge that embody a company standard, but on claude.ai the custom skills you upload yourself are private to your own account by default. On a Team or Enterprise plan you can share your own skill directly with specific colleagues, on Enterprise also with a group, or publish it in the organisation’s library, after which anyone in the organisation can install it. An Owner switches these options on via Organization settings > Plugins & skills, can require a review before publication and can make skills available to the whole organisation.
Shared skills are read-only: colleagues enable and use them, but do not change the content. When the creator updates a shared skill, recipients get that version the next time they use it. You update a published skill by publishing it again, and if a review is required, everyone stays on the approved version until the update is approved.
For a team that wants the same skill used consistently everywhere, this still calls for a fixed owner per skill, even though the marketplace mentioned earlier helps with bundled roll-out on Team and Enterprise plans.
Risks of Claude skills
The risks of Claude skills lie mainly in security and in the lack of central control over the content.
Skills can run scripts in a code execution environment and therefore present a potential attack surface. Anthropic explicitly advises using only skills from trusted sources: skills you have made yourself or that come directly from Anthropic.
A malicious skill from an unknown source can give Claude instructions that differ from what the skill claims to do, potentially leading to data loss or unauthorised access. For organisations that want to use third-party skills, auditing them beforehand is time well spent.
A second risk is less visible but just as real: without a central owner, different versions of the same skill slowly drift apart, resulting in inconsistent output.
My years at major Dutch banks taught me how quickly this happens: without a central owner and periodic review, internal standards slowly drift apart. A simple protocol (one owner per skill, a shared repository, a quarterly review) stops a skill from gradually losing its value.
Who are Claude skills suitable for?
Working with Claude skills suits two groups. The first is individual professionals who regularly do recurring work with their own formats or standards, such as consultants, content creators and analysts. For them, a handful of custom skills works as a personal way of working that Claude applies automatically when relevant.
The second group is organisations that want to standardise their way of working across several employees. A skill for the house style, a skill for the standard contract review, a skill for the in-house reporting format: three examples that add value in a team straight away, because everyone produces the same output without having to explain the standard again in every prompt.
That uniformity is exactly what a team gains once skills are managed centrally, rather than springing up ad hoc for each employee.
If you are just starting with Claude, it makes sense to get the basics right first (chats, projects, prompting) before you move on to skills. Skills do not solve problems you could not already solve in a normal chat; they automate solutions you previously found by hand.
If you jump to skills without that foundation, the result may disappoint you, because the gain from a skill builds on the work that comes before it. The online Claude Course covers how to build that foundation properly and at what point skills actually add value, as part of a broader way of using Claude productively.
Frequently asked questions about Claude skills
Claude skills tie in with other parts of Claude, such as the API and connectors.
Can you use Claude skills through the API?
Yes, you can also use Claude skills through the Claude API, both the ready-made skills and skills you make yourself. You call the built-in skills for PowerPoint, Excel, Word and PDF with their skill_id, and you upload your own skills through the Skills API. Skills run in the container of the code execution tool, so that tool has to be part of your request. A skill you upload via the API is available to all members of your workspace.
Where can you find ready-made Claude skills?
You will find ready-made skills in claude.ai under Customize > Skills, where you can switch example skills on or off on every plan, including Free. Anthropic also publishes open-source skills in its own repository on GitHub. Skills from partners in the Skills Directory are downloaded through the linked repository, usually on GitHub, and are designed to work together with that partner’s MCP connector.
What is the difference between a Claude skill and an MCP connector?
The difference between a Claude skill and an MCP connector is that a connector links Claude to data and systems, while a skill records what Claude does with that data. Anthropic describes MCP as an open standard for connecting AI assistants to the external systems where data lives. A skill is a folder of instructions and, optionally, scripts. Anthropic recommends combining the two, with MCP for the connection and skills for the way of working.
