Agent Skills move to AAIF: inventorying and verifying in-house skills

Checking quotes, formatting meeting minutes, following internal request procedures. Agent Skills is a format for packaging internal procedures like these into a file called SKILL.md and handing them to AI agents. Can a skill built in Claude Code be used as is in another agent, and how will the format change from here? The announcement on October 22, 2026 bears on those assumptions.

On the same day, Anthropic announced that it had donated the Agent Skills standard to the Agentic AI Foundation (AAIF) under the Linux Foundation. AAIF is an organization that Anthropic co-founded with Block and OpenAI, and the Model Context Protocol (MCP) has already been donated to it. On the same day, AAIF also announced on the Linux Foundation site that it had accepted four new projects, including Agent Skills.

Today, we are donating the standard to the Agentic AI Foundation (AAIF) under the Linux Foundation

What moves and what stays the same

According to the announcement, what was donated is the Agent Skills name and logo, the GitHub organization, and the agentskills.io domain. It says that you can continue to use and contribute to the standard as before without joining AAIF.

You can continue to use the standard or contribute to it without joining the AAIF.

The standard will be maintained by engineers from Anthropic, Google, OpenAI, and others in their individual capacities. For the way forward, a roadmap has been proposed that includes the following three items.

• Versioned releases
• A formal proposal process
• Security guidelines for people who write and install skills

The announcement says that since the standard was published in December 2025, more than 70 agent products, including ChatGPT, Cursor, GitHub Copilot, and VS Code, have adopted Agent Skills. One million skills have been published, with nearly 280 million installs.

This donation does not immediately change how existing skills are written. What changes is how the standard is decided, and once versioning begins, you will need to be aware of which version of the format your skills follow.

A refresher on the skill format

In the specification (docs/specification.mdx in the agentskills/agentskills repository), a skill is a folder that contains SKILL.md. The following fields go in the YAML at the top of SKILL.md (the frontmatter).

There are also license and metadata. At startup, an agent reads only the name and description of every skill, and loads the body when it decides to use a skill. The specification recommends keeping the body within 500 lines and splitting detailed material into places such as references/.

Validating Japanese skills with skills-ref

The specification points to the reference library skills-ref for validating skills (its README says it is "for demonstration purposes and not intended for production use"). In the editorial team's test environment (Ubuntu 24.04.5, Python 3.11.17), we installed skills-ref 0.1.0 from the repository's main branch (commit 69ef37e), created several skills containing Japanese, and ran skills-ref validate.

skills-ref validate quote-check
# Valid skill: quote-check

We learned two things.

First, the description limit is counted in characters. We were able to write up to 1,024 characters even in Japanese.

Second, the specification text and the reference tool disagree on how name is handled. The specification says "Lowercase letters, numbers, and hyphens only" in its table and "unicode lowercase alphanumeric characters (a-z, 0-9) and hyphens" in its body text, and the only characters it lists are lowercase English letters, numbers, and hyphens. skills-ref, on the other hand, passed 見積チェック. The skills-ref source (validator.py) says that names can use internationalized characters (Unicode letters) and hyphens, so it intentionally allows Japanese and other scripts.

This discrepancy has also been raised in the repository. There is a proposal to align the specification text with the reference tool (pull request #345) and an issue asking to clarify which one should be followed (#560), but as of October 27, 2026, neither has been resolved.

We did not check this time how each agent product handles Japanese names. Once versioning begins, differences like this may be sorted out. The editorial team recommends keeping the name of skills distributed internally to lowercase English letters, numbers, and hyphens, as the specification text says. Writing Japanese in description and the body gives agents clues for choosing the skill.

What to get in order with internal skills now

What follows is the editorial team's proposal. Getting the following points in order before the format is managed by version will make the transition easier.

1. Make an inventory: List which skills exist in which repositories or shared folders and who maintains them. Also check whether old versions of the same procedure remain in multiple places
2. Standardize the format: Run skills-ref validate in CI to detect mismatches between names and folder names, uppercase letters, and descriptions that are too long. Because it is a reference tool, treat the results as a guide. Note that skills-ref on the main branch reads SKILL.md without specifying an encoding, and there is a report (issue #536, unresolved) that it stops when validating a SKILL.md containing non-ASCII characters (such as Japanese) on Windows. The editorial team tested only on Linux and has not checked on Windows
3. Check the scripts: For skills that include executable code in scripts/, check what it runs and where it connects. When installing skills published externally, read their contents before installing them. The security guidelines are still at the proposal stage and have not been published
4. Do not rely too much on allowed-tools: It is an experimental field in the specification, and support varies by agent. Restrict permissions in the agent product's own settings as well

How to divide roles among instruction files for AI is covered in AGENTS.md / SKILL.md / DESIGN.md. Carrying agent extensions between products is covered in Agent Plugins, and the future of MCP, which was donated to AAIF earlier, in The MCP roadmap.

Scope of verification

On October 23, 2026, we opened and checked Anthropic's announcement "Donating Agent Skills to the Agentic AI Foundation" (October 22, 2026) on anthropic.com. We also checked the specification (docs/specification.mdx) in the GitHub agentskills/agentskills repository (main branch, commit 69ef37e) and the README and source of skills-ref (document research). On October 27, 2026, we opened the agentskills.io specification page, the Linux Foundation announcement, and the AAIF project page, and also confirmed that the main branch was unchanged at 69ef37e. In the editorial team's test environment (Ubuntu 24.04.5, Python 3.11.17), we installed skills-ref 0.1.0 and ran validate and to-prompt on skills created for illustration (hands-on testing). We have not verified loading in each agent product such as Claude Code, ChatGPT, or Cursor, the handling of Japanese names, or script execution.

For design and development that turns internal procedures into a form AI agents can use, contact us through Development, AI, and automation consultations.

References

• Donating Agent Skills to the Agentic AI Foundation（Anthropic, 2026-10-22）
• Agent Skills specification（agentskills/agentskills, GitHub）
• skills-ref（agentskills/agentskills, GitHub）
• Agentic AI Foundation Announces Four New Projects: Agent Skills, Open Responses, OpenEnv, and OpenSandbox（Linux Foundation, 2026-10-22）
• Clarify Unicode support in skill name spec（agentskills/agentskills #345）
• Clarify non-alphanumeric characters in name field（agentskills/agentskills #560）
• skills-ref: validate crashes on Windows for any SKILL.md with non-ASCII characters（agentskills/agentskills #536）