An Agent Skill is a reusable set of instructions and supporting resources that helps a compatible AI agent perform a specific type of task. Instead of explaining the same process every time, you give the agent a skill designed for that work.

A presentation skill, for example, can tell an agent how to open a deck, edit one slide and save the file so it still opens in PowerPoint.

Skills are the fastest-moving part of this ecosystem, so treat this guide as a map rather than a specification.

1. What is an Agent Skill?

Skills are usually folders built around a file called SKILL.md, which holds a short description of what the skill does and when to use it, followed by instructions the agent follows.

Depending on the skill, the folder may also include metadata the agent matches against your request, workflow steps so a multi-stage task repeats in order, templates or reference files such as brand rules, and scripts the agent can run when output has to be exact.

Not every skill includes all of these. Some are little more than a well-written checklist; others bundle code.

2. A practical example: the PDF skill

Anthropic publishes skills in an open repository, including document skills for Word, Excel, PowerPoint and PDF. Its README describes them as the ones that power Claude's document capabilities, released under the Apache 2.0 licence.

The pdf skill is listed in our directory with the procedures its publisher documents: extracting text and tables, merging or splitting files, adding watermarks, filling forms, and running OCR on scanned pages.

A matching request is easy to picture: someone asks an agent to pull the tables out of three supplier PDFs into one file. The agent reads the instructions, follows the extraction step, runs the bundled script where output must be deterministic, and returns a combined file. The person still checks the pages and the numbers.

That describes how the skill is documented to work. It has not been installed or run for this article, and results vary with the agent, model and file.

3. How Agent Skills work

The pattern is similar across the products that support skills. You make a request, such as turning a report into a deck. The agent matches it against the short descriptions of the skills available, then follows that skill's instructions, reading reference files or running a script where needed. You review the result: a draft of a deliverable, not a guarantee.

Behaviour depends on the agent and on how the skill was written.

4. AI tools vs Agent Skills

AI toolAgent skill
What it isA standalone application or platform used directlyAn installable or configurable capability
Where it runsOn its own website or appInside a compatible AI agent
How work startsYou open it and work in itYou ask the agent, and the skill is selected when relevant
What you evaluateFeatures, price, data policyThe publisher, the source, and the permissions it needs

An Agent Skill generally depends on an existing agent environment: it is extra procedure loaded into one, not a substitute.

5. Skills vs prompts, plugins and integrations

These terms get used loosely, and they are not interchangeable.

  • A prompt is an instruction you type into a conversation. It applies to that conversation, and you repeat it when you need it again.
  • A skill is a package of instructions and resources the agent loads by itself when a task matches, and it persists between sessions.
  • A plugin or extension adds capability to one specific application, such as a coding environment.
  • An integration or connector links the agent to another service; the Model Context Protocol is a common way to build those connections.

One capability might involve both: an integration that reaches your files, and a skill that says what to do with them.

6. Which AI agents support skills?

Support is per product rather than per company, and it changes.

  • Claude's products use skills directly. Anthropic's documentation describes pre-built document skills on claude.ai and in the API, plus custom skills you can add in settings or create in Claude Code.
  • Claude Code is a separate agent, a coding tool rather than the Claude chat app. Skills work there too, and many listings in our directory name both.
  • Cursor documents Agent Skills as folders containing a SKILL.md, read from a skills directory in your project.
  • Codex is OpenAI's coding agent, distinct from the ChatGPT app. OpenAI's documentation now covers skills, but availability varies by version and surface.

Two distinctions prevent most of the confusion. Claude and Claude Code are different products from the same company, and a skill that works in one is not automatically available in the other; ChatGPT and Codex work the same way. Our directory records compatibility per agent and lists only documented support.

7. How to find and install skills

  1. Find a skill from an identifiable publisher. A named organisation with a public repository beats an anonymous file share.
  2. Check compatibility against the agent you actually use.
  3. Review the source and permissions. Read the instructions, look at the scripts, note what they access.
  4. Follow the publisher's installation instructions. Environments differ, so there is no universal command to copy.
  5. Test with a low-risk task on a copy of a file you can afford to lose.

8. Are Agent Skills safe?

Skills can contain instructions and executable code. That combination deserves the care you would give a script from the internet, because that is effectively what it can be.

  • Check who published it and where the source lives. Our skill listings record the publisher, repository, licence and source.
  • Question the permissions. Files, accounts, network access and outside services all matter.
  • Avoid untrusted scripts. An instruction to run unknown code with broad access to your machine is a reason to stop.
  • Give the minimum access it needs, and keep sensitive material out of reach until you trust it.

One caveat: source verification is not a security audit. It tells you where the code came from, not that the code is safe, and it is not an endorsement.

Start with one skill

The best way to understand skills is to use one that does something small and checkable. Browse the Agent Skills directory, filter by your agent, and start with a skill that reads data before you choose one that can change or delete it.