Claude Skills: What They Are and How to Use Them

Claude has always been able to help with a wide range of tasks. The problem is that without some form of persistent instruction, every session starts from scratch. You re-explain your preferred format, remind it which template to use, describe the workflow again. Skills fix that.

A Claude skill is a folder of instructions, scripts, and optional resources that Claude loads when it recognizes a matching task. Once a skill exists, Claude knows how to perform that task the same way every time, without you spelling it out. The official definition describes skills as teaching “Claude to complete specific tasks in a repeatable way,” and that is the clearest summary of what they actually do.

Why skills exist

The core problem skills solve is repetition. If you analyze data the same way each week, create documents to a specific brand standard, or automate a personal task that follows a fixed set of steps, you should not have to re-explain that process every time. A skill captures the process once and makes it available on demand.

This also matters for quality. When Claude is working from a skill rather than improvised instructions, the output is more consistent. The skill can include not just what to do, but how to structure the result, which tone to use, which checks to run, and what to do when something unexpected comes up.

How Claude picks and loads a skill

When you send a request, Claude reviews the skills available in your environment and identifies which ones are relevant to what you asked. It then loads only those skills. This approach is called progressive disclosure: Claude brings in what it needs and leaves everything else out.

The reason progressive disclosure matters is context. Claude’s context window is finite. Loading every skill every time would crowd out the actual content of your task. By loading selectively, Claude keeps its context focused and performs better.

You do not have to name the skill explicitly. Claude infers which skills apply based on your request. That said, in environments like Claude Cowork, skills are surfaced more visibly, so you can see what is active and trigger specific ones intentionally.

The three types of Claude skills

Skills come in three categories, each with a different source and a different purpose.

Anthropic skills are built and maintained by Anthropic. They load automatically when relevant, without any setup on your part. The clearest examples are the document creation skills: if you ask Claude to produce an Excel spreadsheet, a Word document, a PowerPoint presentation, or a PDF, Anthropic skills handle the formatting and structure automatically. You do not configure these. They are just there.

Custom skills are created by users or organizations. A marketing team might build a skill that applies their brand voice to every piece of copy. A solo operator might build a skill that runs their weekly review in a fixed format. Custom skills are defined by you, scoped to what you need, and available across sessions once they exist. Creating one is covered in more detail below.

Partner skills come from third-party services that have built Claude integrations. These work alongside MCP connectors (more on connectors in a moment) from partners like Notion, Figma, and Atlassian. A partner skill teaches Claude how to interact with that service in a structured way, rather than relying on improvised instructions each time.

The full breakdown of skill types, including how they interact with connectors and the broader Claude ecosystem, is covered in the Claude Cowork skills guide.

How to start using built-in skills today

If you are on any Claude plan with code execution enabled, Anthropic skills are already active. You do not enable them manually. The next time you ask Claude to create a formatted Word document, an Excel file, or a PDF, you will see the skill in action: Claude produces a structured, downloadable file rather than plain text output.

The most direct way to explore skills more fully is through Claude Cowork. Cowork is the environment where skills are most visible as a feature. You can see which skills are available, invoke them by name, and build workflows around them. If you are new to Cowork, the beginner course at /learn walks through skills as part of a broader introduction to how Cowork is structured.

If you want to go deeper on how skills, connectors, and scheduled tasks fit together as a system, the Cowork connectors, skills, and scheduled tasks overview is the right place to start.

Creating your own custom skill

Building a custom skill does not require a software background, but it does require understanding the basic structure. A skill is a folder. Inside that folder is at minimum one instructions file, typically a markdown file that tells Claude what to do and how to do it. You can also include scripts (code Claude can run when executing the skill) and resources (templates, reference files, data Claude should consult).

The instructions file is the core of the skill. It should describe the task clearly, specify the expected output format, and include any constraints or rules Claude should follow. The more specific and honest the instructions, the more reliably Claude performs.

One thing worth knowing: skills follow an open standard called the Agent Skills specification, published at agentskills.io. This means skill files are not locked to Claude. A skill you write today could, in principle, work with other AI systems that adopt the same specification. That is a meaningful architectural choice. It keeps your investment in building skills portable.

If you want a full tutorial on building and testing custom skills, the Claude Cowork skills guide goes deeper than this overview.

Skills vs. connectors vs. plugins

These three terms come up together often enough that a brief clarifier is worth including.

A skill teaches Claude how to perform a task. It is about behavior and process.

A connector links Claude to an external service so it can read and write data there. Connectors use the Model Context Protocol (MCP). If you want Claude to pull content from Notion, post to a project management tool, or interact with Figma, you need a connector. The Claude connectors guide covers setup in detail, and the MCP explainer is useful background if the protocol is new to you.

A plugin extends what Claude can do within a specific environment, like adding a tool or integration to your Cowork setup. Plugins and skills can work alongside each other, but they are not the same thing. The Claude Cowork plugins guide explains the distinction with examples.

In practice, you often use all three together. A skill defines the workflow. A connector pulls in the data the workflow needs. A plugin adds a capability Claude would not otherwise have in that environment.

The code execution requirement

Skills require code execution to be enabled. This is because many skills run scripts as part of their process. If code execution is off, Claude cannot run those scripts, and skills that depend on them will not work correctly.

Check your Claude settings and confirm code execution is active before expecting skills to behave as documented. This applies across all plan tiers, including free.

What skills are not

Skills are not a replacement for clear instructions in the moment. If you are doing something one-off, a prompt is faster and more appropriate. Skills pay off when a task recurs: the same report format, the same analysis workflow, the same content structure. That is where the upfront investment in defining the skill returns value.

Skills also do not make Claude infallible. A poorly written skill produces poorly structured output, consistently. The quality of a custom skill depends on the quality of the instructions you write for it. Start narrow, test with a specific task, and refine from there.

The beginner course at /learn walks through exactly this process in the context of Cowork, which is the most practical environment for putting skills to work from day one.