Agent Skills

Anthropic shipped Agent Skills in October as a Claude feature: a skill is a folder with a SKILL.md file — a few lines of YAML naming the task, then plain Markdown instructions, plus any scripts the job needs. The agent reads only the one-line description until a task matches, then pulls in the full skill. No SDK, no new syntax, nothing to compile.

Claude Skills are awesome, maybe a bigger deal than MCP. — Simon Willison

The lightness is the whole story. Because a skill is just Markdown in a folder, a rival agent can read it with no permission and no porting. OpenAI's Codex merged exactly that — a feature literally named 'experimental support for skills.md' — on December 2nd, six days before Anthropic got around to declaring the format an open standard. A competitor adopted the spec before its author called it one.

By spring, dozens of tools read the same folders: OpenAI's Codex and ChatGPT, Google's Gemini CLI, Microsoft's Copilot, Cursor, JetBrains, AWS. The same plain file that teaches Claude to fill out a spreadsheet teaches a competitor's agent the same trick, unedited.

This is Anthropic's second swing at owning the plumbing under every agent — the first, the Model Context Protocol, it just handed to the Linux Foundation. One analyst's standing objection is that single vendors rarely set standards alone. The Codex merge is the counter-evidence: the thing was small and useful enough that rivals copied it on their own, before anyone was asked to.