What are AI Agent Skills?

Agent Skills are a lightweight, open format for giving AI agents specialised knowledge and workflows. A skill is simply a folder that contains a SKILL.md file with a name, a description, and instructions — plus any scripts or reference files the agent might need. When a task matches what a skill does, the agent loads it on demand.

This guide explains what an Agent Skill is, why the format exists, how discovery and activation work, and how skills differ from prompts, tools and traditional plugins. If you want the file format itself in depth, read the SKILL.md guide.

What is an Agent Skill?

An Agent Skill is a portable package of expertise. Instead of hard-coding behaviour into an application, you describe it in a folder the agent can read. The skill tells the agent how to do something — for example, run a compliance review, format a document a specific way, or operate a niche tool — and can include supporting scripts and references.

The minimum skill is remarkably small:

my-skill/
└── SKILL.md

A SKILL.md holds metadata (at least a name and description) and the instructions themselves. Larger skills add scripts/, references/, and assets/ directories for code, documentation and templates.

Why Agent Skills exist

Modern agents are capable, but they often lack the specific context needed to do real work reliably. A general model doesn't know your team's review process, your data pipeline conventions, or the quirks of an internal tool. Skills solve this by packaging that knowledge into version-controlled folders that any compatible agent can load.

The benefits are practical:

• Domain expertise — capture specialised knowledge as reusable instructions.
• Repeatable workflows — turn multi-step tasks into consistent, auditable procedures.
• Cross-product reuse — write a skill once and use it across skills-compatible agents.

The format is open, originally developed by Anthropic and now contributed to by a broad ecosystem.

How skills are discovered

Agents use progressive disclosure to keep their context lean. At startup, an agent loads only the name and description of each available skill — just enough to know when a skill might be relevant. Full instructions never enter context unless they are needed.

This matters because an agent might have access to dozens or hundreds of skills. Loading every instruction up front would waste context and degrade performance.

How a skill is activated

Activation happens in three stages:

1. Discovery — the agent reads each skill's name and description.
2. Activation — when a task matches a description, the agent reads the full SKILL.md into context.
3. Execution — the agent follows the instructions, optionally running bundled scripts or loading referenced files.

Because full instructions load only when a task calls for them, agents can keep many skills on hand with a small context footprint. To learn how to write descriptions that activate reliably, see the guide to creating an Agent Skill.

What a SKILL.md file contains

A SKILL.md typically includes:

• Name — a short identifier for the skill.
• Description — what the skill does and when to use it (this drives activation).
• Instructions — the main body, written as natural-language steps.
• Optional references to scripts, reference files, and assets.

For the full structure and a worked example, see SKILL.md: the file behind Agent Skills.

Instructions vs scripts vs references

These three layers serve different purposes:

• Instructions are natural-language guidance the model reads. They carry most of a skill's value.
• Scripts are executable code the agent can run when needed — for example, a validator or a data transform.
• References are supporting documents the agent loads selectively, so large material doesn't bloat context.

A good skill leads with clear instructions and adds scripts and references only when they genuinely help. More detail is in the security guide, which covers what to check before installing a skill.

Skills vs prompts

A prompt is a one-off instruction for a single conversation. A skill is durable, reusable expertise that an agent can load across many tasks and sessions. Prompts live in chat history; skills live in version-controlled folders that travel with your workflow.

Skills vs tools

A tool is a capability the agent can call — "fetch this URL", "run this query". Skills are higher-level: they package know-how that may orchestrate tools, references and judgement. Skills and tools are complementary; an agent can use both at once.

Where Agent Skills are used

Skills are now documented across major agent ecosystems. See the dedicated guides for specifics:

• Claude Skills
• Codex Skills
• ChatGPT Skills

A growing number of coding agents and runtimes also support the format, so a skill written once can move between them.

Common Agent Skill use cases

Typical use cases include:

• Codifying a team's code-review or documentation standard.
• Automating a compliance or QA checklist.
• Operating a niche internal tool via scripts.
• Formatting and producing domain-specific documents.
• Teaching an agent a proprietary workflow.

Benefits and limitations

Skills make expertise portable, repeatable and auditable. They also have honest limits: a skill is only as good as its instructions, and because skills can run scripts they deserve the same scrutiny as any other code (see Agent Skill security). Skills are a way to package knowledge — not a guarantee of correctness.

Frequently asked questions

Do I need to write code to create a skill? No. Many useful skills are pure instructions in SKILL.md. Scripts are optional.

Are skills tied to one vendor? No. The format is open and supported across multiple ecosystems, with an open specification.

How is a skill different from a system prompt? A system prompt configures one session; a skill is reusable, selectively loaded expertise that persists across sessions and products.

Can skills work together? Yes. Agents can hold many skills and activate the relevant ones per task.