A short definition
An agent skill is a small, self-contained package of instructions that teaches an AI agent how to perform a particular job. Instead of explaining the same workflow to your assistant every time, you bundle the steps, conventions, and gotchas once. The agent then loads that skill when it is relevant and follows it.
Think of a skill the way you would think of an onboarding doc for a new teammate. It does not give the teammate new hands or new accounts. It tells them, in your house style, exactly how a task should be done here: what order to do things in, what to double-check, and how to format the result.
What lives inside a skill
At the center of every skill is a file called SKILL.md. It holds a name, a description that tells the agent when to reach for the skill, and a body of instructions written in plain prose. A skill may also ship supporting files: reference documents the agent can read on demand, example templates, or small helper scripts. The point is that everything needed to run the workflow travels together in one folder.
Because a skill is just text and files, it is easy to read, edit, and version with the rest of your project. You can open it, see exactly what it tells the agent to do, and change a line you disagree with.
Why skills matter
Without skills, useful know-how stays trapped in chat history or in one person's head. With skills, that know-how becomes a portable asset. The same skill can be shared across a team, dropped into a new project, or published for others to install.
Skills also make agent behavior more predictable. A good skill encodes the boring-but-important details: the naming convention, the validation step people forget, the format the downstream tool expects. The agent stops improvising and starts following a known-good path.
Where you might use one
Skills show up across many kinds of work. On the development side you will find skills for code review, refactoring, and writing tests. There are skills for documentation, for data cleanup, for design handoff, and for research. A helpful way to start is to browse the Skills directory and narrow it by category or Agent platform.
If you are new, pick one narrow task you repeat often and look for a skill that targets exactly that. A focused skill almost always outperforms a broad, do-everything one.
How skills relate to the rest of the toolkit
It helps to keep two ideas separate. A skill is procedural knowledge — the recipe. The agent still needs hands to act: a code editor, a terminal, a browser, or external tools. Those capabilities often come from elsewhere, such as MCP servers. A skill tells the agent how to use what it already has; it does not, by itself, grant new powers.
This separation is why skills stay small and readable. They lean on the agent's existing abilities and add the missing ingredient: judgment about how to apply them in your context.
Getting started
The fastest way to understand skills is to install one and watch it work. Find a skill that matches a task you do weekly, read its SKILL.md so you know what it will do, then add it to your agent. From there you can tweak the instructions to fit your team's conventions.
How skills differ from prompts
A natural question is how a skill differs from simply typing a good prompt. The difference is durability and reach. A prompt lives in one conversation and disappears when it ends. A skill lives outside any single chat: it has a name, a trigger description, and supporting files, and the agent loads it automatically when it is relevant. A prompt is a sentence you type once; a skill is a saved procedure that runs again and again, the same way every time. The deeper comparison lives in agent skills vs prompts.
What makes a skill good
Not all skills are equal. The strong ones share a few traits. Their trigger description is specific, so the agent knows exactly when to use them. They stay focused on a single workflow rather than trying to do everything. They are well documented, with clear steps and a stated output. And they are honest about anything they touch — scripts, network calls, or credentials. When you browse a curated directory, these are the qualities the listings are screened for, and they are the same qualities you should check yourself before installing anything.