What Is an AI Agent Skill? It's a Folder, Not a Prompt

For most of AI's history, getting a model to do something specific meant writing a long, clever prompt — a wall of instructions you paste in every time. That changed in October 2026. The unit of reuse stopped being a paragraph and became a file format: a skill is a folder an agent can carry, and the ecosystem has now agreed on what goes inside it.

What a skill actually is

A skill is not a prompt snippet. It is a folder containing a SKILL.md file — two required metadata fields (name and description) plus the instructions that tell an agent how to do one task. The folder can also carry scripts/, references/, assets/, or anything else the task needs.

That is the entire specification. No SDK, no runtime, no registry required. Which is exactly why it spread.

How the agent loads it: three stages, not one

The loading mechanism is the interesting part, and it has a name: progressive disclosure.

1. Discovery. At startup the agent reads only each skill's name and description — not the body. Just enough to judge whether the skill might be relevant.
2. Activation. When a task matches that description, the full SKILL.md gets pulled into context.
3. Execution. The agent follows the instructions, running bundled code or opening referenced files only if the task calls for it.

The consequence is structural: an agent can hold hundreds of skills at a context cost of a few lines each. Skills don't compete for context the way tools and documents do — because most of them are never loaded.

Why this beat prompts

• Portable. A folder travels. Copy it between products, commit it to git, version it.
• Addressable. The description is a routing key. The agent decides relevance from it before spending tokens on the body.
• Auditable. The procedure is a file you can read and diff — unlike a system prompt assembled at runtime.

The standard arrived

On 22 October 2026 Anthropic donated the Agent Skills name, logo, GitHub organisation, and the agentskills.io domain to the Agentic AI Foundation, under the Linux Foundation — the same home as MCP. The published numbers: more than 70 agent platforms have integrated the standard since its December 2025 release, 1 million skills have been published, and they have been installed nearly 280 million times. Named adopters include ChatGPT, Cursor, GitHub Copilot, and VS Code.

Two things are worth noticing. The format was invented inside Claude Code and then handed to a neutral foundation — standardisation as a distribution strategy, not a concession. And the adoption numbers describe an install base, not a quality bar.

Why this matters for you

If you are building agents, stop treating every instruction as a one-off prompt. Package repeated know-how as a folder with a description sharp enough to route on — the description does more work than the body.

But the hard problem has not moved. Research on skill transfer found that task-level skills often perform worse than having no memory at all, and that plain-text descriptions transfer better than code. A million published skills is evidence of supply, not of usefulness.

Skills sit between your framework and your tools — part of the architecture that makes an agent more than a chatbot. How frameworks like LangGraph, CrewAI, and the OpenAI Agents SDK fit together is compared side by side in our agent framework comparison.

how do AI agents work — return to the complete AI agent architecture guide.