Let us begin with a plain definition.

According to Anthropic’s help centre, custom Skills let you enhance Claude with specialised knowledge and workflows specific to your organisation or your personal work style. A skill is a folder containing a main instruction file, usually called SKILL.md, and optionally other files such as templates, reference notes or small programs. The instruction file explains what the skill is for and how Claude should do the task.

When you create a skill, Anthropic describes it as giving Claude a playbook it can reference whenever you need a particular type of help, such as generating reports in your company’s format, cleaning and using data the way you normally do, or analysing customer data your way. Skills can be as simple as a few lines of instructions or as complex as multi-file packages with executable code.

Skills solve a very ordinary problem: repeating yourself.

Anthropic’s guidance says that a skill pays off when Claude does a task for you repeatedly and you want it done the same way every time. Without skills, you might paste the same long instructions into every new chat, or hope Claude guesses your preferred format. That is slow, and the results can drift from one chat to the next.

Skills also help with tasks that need more than good wording. A skill can carry a template, a checklist, reference material or a small program, so that Claude has everything it needs in one tidy package. This makes results more consistent, because the same instructions are followed each time.

flowchart LR
  A["Without a skill"] --> B["Explain your method in every chat"]
  B --> C["Results vary from chat to chat"]
  D["With a skill"] --> E["Write the playbook once"]
  E --> F["Claude loads it when the task appears"]
  F --> G["More consistent results"]
        

Claude has several ways to give it context, and skills are easy to confuse with the others.

flowchart TD
  A["What do you want to give Claude?"] --> B{"Type of help"}
  B -- "One-off instructions" --> C["Put it in your prompt"]
  B -- "Background for one project" --> D["Use project instructions and files"]
  B -- "Facts about you" --> E["Memory"]
  B -- "Access to an app or data" --> F["Connector"]
  B -- "A repeatable way of doing a task" --> G["Skill"]
        

A skill has a simple structure, and knowing it makes everything else clearer.

Every skill is a folder. Inside is one required file, SKILL.md. At the top of that file sit a few lines of information about the skill, most importantly its name and a description of what it does and when to use it. Below that is the main body of instructions, written in plain language. Anthropic’s guidance says the folder name must match the name field.

Beyond SKILL.md, everything else is optional. A skill may include extra reference files, templates, example inputs and outputs, or scripts, which are small programs Claude can run. Anthropic’s guidance says that Claude reads all of SKILL.md every time the skill loads, so long material needed only some of the time is better kept in separate files that SKILL.md points to. It also suggests mentioning each file at the step where Claude should use it.

flowchart TD
  A["Skill folder"] --> B["SKILL.md: name, description, instructions"]
  A --> C["Reference files, optional"]
  A --> D["Templates and examples, optional"]
  A --> E["Scripts, optional"]
  B -.-> C
  B -.-> D
  B -.-> E
        

This is the clever idea that lets Claude carry many skills without being overwhelmed.

Claude can only hold a limited amount of material in view at one time, a space called the context window. If every skill were loaded in full, that space would fill up quickly. Anthropic’s design avoids this with what it calls progressive disclosure: information is revealed in layers, only as needed.

In the first layer, Claude sees only each skill’s name and description, which is a very small amount of text per skill. This is enough to know which skills exist and roughly when each applies. If a task seems to match, Claude moves to the second layer and reads the full body of SKILL.md. If the instructions point to extra files or scripts, Claude reaches the third layer, opening them only when a step calls for them.

flowchart LR
  A["Your request"] --> B["Claude scans skill names and descriptions"]
  B --> C{"Does a skill match?"}
  C -- "No" --> D["Answer normally"]
  C -- "Yes" --> E["Load that skill's SKILL.md"]
  E --> F{"Does a step need extra files?"}
  F -- "Yes" --> G["Open only those files"]
  F -- "No" --> H["Follow the instructions"]
  G --> H
        

Examples show how wide the range is.

Anthropic provides some skills itself, and Claude’s own file creation uses what the help centre calls document skills, which bundle instructions and helper programs for producing documents, spreadsheets and presentations. Anthropic also maintains a public collection of example skills that you can use as templates. Beyond these, individuals and companies create their own.

You usually do not have to say “use this skill”. Claude chooses.

Anthropic’s guidance explains that Claude uses descriptions to decide when to invoke a skill. When you make a request, Claude compares it with the descriptions of the skills that are turned on, and loads the ones that seem relevant. This means that the quality of your description affects how reliably the skill triggers. You can also name a skill in your message to nudge Claude toward it.

Several skills can work together on one task. The help centre notes that although skills cannot explicitly reference each other, Claude can use multiple skills together automatically, which it describes as one of the most powerful parts of the feature. For example, a brand-style skill and a report-writing skill might both apply to a single document.

Availability depends on your plan and where you use Claude.

Skills work across several Claude surfaces. Coverage says they are supported in the Claude apps, in Claude Code, in the developer tools and in the API. Coverage of Anthropic’s announcements says that skills are included at no additional cost on paid plans, and one guide notes that skills are not available on the free plan. Anthropic’s help article on creating skills mentions the Skills area under Settings, in the Capabilities section, where you can switch skills on and off.

The easiest way to make a skill requires no file editing at all.

Anthropic’s help centre describes two paths to creating skills. In the first, you write the files yourself for full control. In the second, you create a skill with Claude through conversation. For this, you turn on a built-in skill called skill-creator in the Skills section of your Settings. This skill gives Claude the ability to build properly formatted skills for you.

Once it is on, you start a chat and describe the skill you want, for example a skill for quarterly business reviews or one that knows how to analyse customer feedback. Claude asks about your process and generates the folder structure and instructions. The help centre advises describing your process in your own words, so that Claude can structure it into proper instructions.

If you want full control, you can build the folder by hand.

Anthropic’s help centre explains how to create, structure and test your own skills. You write a SKILL.md file with a name, a description and clear instructions, add any supporting files, and then package the folder as a ZIP file for upload. The guidance stresses one detail: the ZIP should contain the skill folder as its root, rather than having files placed directly at the top level of the archive. If you use another tool to create the ZIP, compress the skill folder itself, not the files inside it.

A few principles from Anthropic’s guidance make skills far more reliable.

Skills are designed to be shared and, increasingly, to be portable.

Coverage of Anthropic’s December 2025 enterprise update says that administrators on Team and Enterprise plans can manage skills from a central location, provisioning them across the organisation and controlling which workflows are available, while individuals can still add their own. Anthropic also described a directory of skills built by partner companies.

Anthropic has also published the skill format as an open standard, documented at agentskills.io, and its help centre encourages following that specification so that skills you create can work across platforms that adopt it. Coverage says that other AI tools have adopted the same folder-and-SKILL.md approach. The practical benefit is that a well-written skill is not locked into one product.

A skill is instructions, and sometimes code, so it deserves the same caution as any software.

Because a skill can tell Claude what to do and can include programs for Claude to run, a careless or malicious skill could cause harm. For example, it could instruct Claude to behave in unexpected ways or to handle your files in ways you did not intend. Anthropic’s guidance, as reflected in coverage of its documentation, stresses using skills from trusted sources and reviewing them before you turn them on.

These errors appear often, and each has an easy fix.

Here is the whole picture in one place.

A folder with a SKILL.md instruction file, and optional supporting files, that teaches Claude a specific workflow or way of doing a task. Claude loads it when a matching task appears.

A prompt applies to one request. A skill is stored and reused, so the same method is available across many chats without retyping it.

Usually not. Claude reads each enabled skill’s description and decides when it applies. You can also name a skill in your message to nudge Claude toward it.

It is the layered loading design: Claude sees only a skill’s name and description at first, loads the full instructions when relevant, and opens extra files only when a step needs them. This keeps the context window uncluttered.

Coverage says skills are included at no extra cost on paid plans, and one guide says they are not available on the free plan. Team and Enterprise plans can have centrally managed skills. Check the help centre for your plan.

Turn on the built-in skill-creator in the Skills section of your Settings, then describe your process in a chat. Claude asks questions and builds a properly structured skill for you.

Compress the skill folder into a ZIP so the folder is the root of the archive, and then upload it in your Skills settings. Check that every entry starts with the folder name.

Yes. Skills cannot explicitly reference each other, but Claude can use several together automatically when a task calls for it, which is why focused skills compose well.

They are as safe as their source. Use skills from trusted authors, read the instructions and any scripts before enabling them, never include secrets in a skill and follow your organisation’s policies.

Anthropic’s help centre has articles on creating skills, and Anthropic maintains a public collection of example skills. The open specification is documented at agentskills.io. Because the feature is evolving, confirm details in the current articles.

A skill is a reusable playbook for Claude: a folder with a SKILL.md file holding a name, a description and instructions, plus optional templates, references and scripts. Claude keeps only the names and descriptions in view, loads the full instructions when a task matches, and opens extra files only when a step needs them, a design called progressive disclosure that lets many skills coexist without crowding its working space. Skills differ from prompts, projects, memory and connectors because they capture a method that repeats across chats and people. You can turn skills on in your Settings, create them by talking to Claude through the built-in skill-creator, or write and package them yourself as a ZIP with the folder as its root. The best skills are focused, have clear descriptions, include examples and are tested step by step. Skills can be shared, managed by administrators on work plans and follow an open standard, but because they can contain instructions and programs, you should use only trusted sources and review them first. As the feature is still evolving, check the current help articles.