What Are AI Agent Skills, Really? The Universal Format That Took Over AI Tooling in Six Months

117,000 GitHub stars. 34,000 community-contributed skills. 18+ AI agents supporting it. The SKILL.md format went from an obscure Anthropic feature to the de facto industry standard in six months. Here is what it actually is and why every developer working with AI agents needs to understand it.

In October 2025, Anthropic released a small new feature inside Claude called Agent Skills. The original goal was modest: help Claude perform better on a few specific task types like creating spreadsheets and working with PDFs. The mechanism was equally modest — a folder containing a markdown file with some YAML metadata at the top.

Nobody at Anthropic predicted what happened next.

From Feature to Standard in Record Time

Within weeks, developers realised the SKILL.md format Anthropic had invented was not just useful for Claude — it could be a universal way to give any AI agent specialised capabilities. The simplicity was the genius. There was no protocol to implement, no API to integrate, no infrastructure to deploy. A folder. A markdown file. That was the entire spec.

By December 2025, Anthropic had published the specification as an open standard at agentskills.io. By January 2026, Cursor had added native support, and Vercel launched skills.sh — a community marketplace that hit 34,000 skills within months. By March 2026, OpenAI Codex CLI, Google Gemini CLI, Microsoft GitHub Copilot, OpenClaw, Hermes Agent and dozens of other tools all supported the same standard.

What a Skill Actually Is

A skill is a folder. Inside that folder, there is a file called SKILL.md. The file has YAML frontmatter at the top with at least two required fields — a name and a description — followed by markdown content explaining what the skill does and when the agent should use it.

That is the entire specification. There is no protocol, no API integration, no installation process beyond putting the folder in the right place. The AI agent reads the metadata at startup, sees there is a skill available, and when your request matches the description, the agent loads the full instructions and follows them.

The Problem It Solves

The problem skills solve is more profound than it appears. AI agents are limited by their context window — the total amount of text they can hold in active memory at once. If you load thousands of tokens of specialised instructions into every conversation, the agent performs worse on everything else. But specialised knowledge is exactly what makes AI agents genuinely useful.

Skills solve this with progressive disclosure. At startup, the agent reads only the brief metadata (around 100 tokens per skill). The full instructions for any skill load only when that specific skill is triggered by a user request. This means you can install 50 skills without paying a 50-skill context tax on every conversation — only the skills that actually match your current request load their full content.

Why Developers Adopted It So Quickly

The single property that drove adoption faster than any other agent technology: skills are agent-agnostic. The same folder works in Claude Code, Cursor, OpenAI Codex CLI, Google Gemini CLI, GitHub Copilot, OpenClaw and any other tool that adopts the standard, without modification. Developers do not have to bet on a single AI vendor. Teams can switch tools without rewriting their accumulated skills library. Organisations can build internal skills that work across whatever AI tools their team chooses to use.

This is the npm moment for AI capabilities — and the entire industry has rallied around it.