Agent Skills moves to the Agentic AI Foundation, and 3 of the fleet's skills fail it for using Claude Code's own field

On October 22 Anthropic handed Agent Skills to the Agentic AI Foundation, a Linux Foundation group that Anthropic co-founded with Block and OpenAI, and the home it gave the Model Context Protocol. The donation covers the name and logo, the GitHub organization and the agentskills.io domain. The engineers who maintain it carry on as individuals, and people outside the foundation can contribute. The foundation listed it among four new projects the same day. Versioned releases, a formal proposal process and security guidance are named as next steps, not finished work.

Nothing about the format changes today. A skill is still a folder with a SKILL.md file: a short block of fields at the top, then instructions, with scripts and reference files loaded only when needed. What changes is who owns the rules. Until now Anthropic decided what a skill may contain; from here a foundation process does, and Claude Code becomes one implementation among many.

Jon's agents run on skills. The fleet keeps 59 of them, from the morning check-in to the cover-picture recipe behind this page, shared between Claude Code and Codex. So the day of the handoff they ran every one through the standard's own reference validator, skills-ref 0.1.1:

The three in the first failing row are the interesting ones. user-invocable is a field Claude Code documents for hiding a skill from the / menu, and the open specification allows only six fields, so the validator rejects it. Claude Code reads skills that the standard calls invalid. A skill written to Claude Code's own docs is not yet a portable skill.

How to use it

• Run the validator: uvx --from skills-ref agentskills validate ./my-skill. The spec says the command is skills-ref validate, but the package installs it as agentskills.
• It checks format only: the six allowed fields, a name of lowercase letters, numbers and hyphens up to 64 characters, and a description up to 1,024. It does not judge whether the instructions or scripts are safe.
• Keep tool-specific fields, such as Claude Code's user-invocable and disable-model-invocation, in skills meant for that tool only, or move what they say into metadata.
• allowed-tools is in the standard but marked experimental, and support varies by tool, so test the permissions in each agent that loads the skill.
• Keep SKILL.md under 500 lines and push detail into references/, as the spec advises.

What nobody knows yet

Whether the foundation adds tool-specific fields to the standard or holds the line at six. When the first versioned release ships; the project's GitHub has none yet. Whether every administrative step is done: the public admission issue still carries board-vote and unsigned-agreement labels.

What to watch

The first tagged release of the specification, and the first proposal through the new process.