Agent Skills Moves to the Linux Foundation's AAIF: What Changed Hands, and a SKILL.md Spec Check

On October 22, 2026, Anthropic donated the Agent Skills standard to the Agentic AI Foundation (AAIF), which sits under the Linux Foundation. The name and logo, the GitHub organization, and the agentskills.io domain changed hands. The SKILL.md format itself did not change, so there is nothing to fix in your code today. What changes is who edits the spec and how. Anthropic's proposed first roadmap covers versioned releases, a formal proposal process, and security guidance for people who write and install skills.

Once a format leaves a single vendor, sticking to the spec matters more. We checked the four SKILL.md files in the Kortress repository, which builds this site, against the published spec. Two of them could not be read by a standard YAML parser. The cause was one colon: an unquoted Trigger: inside the description.

This post covers what moved and what stayed, how this differs from the MCP donation, our own check, a third-party security scan, and early community reaction.

What moved and what stayed

Agent Skills grew out of Anthropic's work on Claude Code. A skill is a folder with a SKILL.md file, plus any scripts or reference material it needs. Anthropic launched skills in October 2025 and published them as an open standard, with its own GitHub organization and domain, that December. According to Anthropic, more than 70 agent platforms have since integrated the standard, including ChatGPT, Cursor, GitHub Copilot, and VS Code, and about 1 million skills have been published with nearly 280 million installs. The platform count is Anthropic's own; we found no independent tally. The skill and install figures match what Vercel reported on September 25 for its skills.sh registry ("one million agent skills", "nearly 280 million installs"), and Vercel notes that installs are aggregate registry counters, not unique people. Anthropic's post does not say where its figures come from.

Here is the ownership change as we read it from the announcement.

One thing the table leaves out: the announcement did not touch the spec text. The spec at agentskills.io still has two required fields (name, description) and four optional ones (license, compatibility, metadata, and the experimental allowed-tools). It has no top-level version field; the spec only shows version tucked inside metadata in an example. Versioned releases are a proposal for now.

How this differs from the MCP donation

Anthropic, Block, and OpenAI co-founded AAIF on December 9, 2025, and Anthropic donated the Model Context Protocol (MCP) as a founding project. This is the same company handing a second standard to the same foundation, so we lined up the two announcements side by side.

With MCP, Anthropic said the governance model would stay the same. This time it named the maintainers and a roadmap up front. The scale figures use different units, so they cannot tell you which standard is bigger: MCP counted servers and SDK downloads, while Agent Skills counts skills and installs.

AAIF's own announcement says the foundation now hosts 10 projects, up from three at launch. OpenAI donating Open Responses on the same day stands out: two competitors are now parking their interface specs in one neutral place.

Checking four SKILL.md files in this repository

A neutral standard only works if different agents read the same skill the same way, and that starts with files that follow the spec. We checked the four SKILL.md files in the Kortress repository against agentskills.io: name format (lowercase letters, digits, hyphens, no leading, trailing, or doubled hyphens, at most 64 characters, matching the folder name), description length (at most 1,024 characters), the recommended 500-line limit, and whether the frontmatter parses as standard YAML. Parsing used Python's PyYAML (yaml.safe_load).

All four pass the name, length, and line-count checks. The two failures stop for the same reason. The parser error was mapping values are not allowed here, pointing at Trigger: inside the description. In YAML, a colon followed by a space inside an unquoted value reads as the start of a new key. The two passing files write the description as a >- block scalar, which sidesteps the problem.

Here is the failing description line next to the fixed one.

# Before (fails to parse)
description: Kortress 계열 프로젝트(…)의 기술 블로그 글 작성 규범. … 쓴다. Trigger: "블로그 글", "실험 노트", …

# After (parses) — the whole value wrapped in single quotes
description: 'Kortress 계열 프로젝트(…)의 기술 블로그 글 작성 규범. … 쓴다. Trigger: "블로그 글", "실험 노트", …'

The fixed version parses with the same PyYAML, and the description is still 206 characters. Yet both broken skills showed up normally in the skill list of the Claude Code session used to write this post. One tool read the files leniently while a standard parser rejected them. As the format spreads across vendors, a file one tool quietly accepts could be skipped entirely by another.

This is worth checking if you write skill descriptions in a language other than English. Labels like "Trigger:" or "Note:" followed by a space trigger the same failure. Quoting the whole value or using a block scalar such as >- avoids it.

A security scan of 9,930 public skills

One third-party measurement shows why security guidance made the first roadmap. The Agent Forge published a report (v1.1) on October 19, 2026 that collected 9,930 unique SKILL.md files from 277 public GitHub repositories and ran 34 text rules over them. No model was used, nothing was executed, and every flagged file was reviewed by hand. The company offers the same scanner for free and has open-sourced it, which the report itself lists as a conflict of interest.

The report says it found no signs of malice in the flagged files. Because the scan is rule-based, files that slip past the rules are not in these numbers. Common medium-severity patterns were turning off permission prompts (28 files), editing standing agent files (27), instructions hidden in HTML comments (16), and reading a .env file (16). 0.33% is small, but a skill is a set of instructions the agent follows as written, so one bad install can do real damage. That looks like the reason security guidance is on the roadmap from day one.

Early community reaction

The Anthropic announcement was posted to Hacker News, but at the time of writing it had 5 points and no comments. There is no real debate there yet.

Simon Willison wrote on his blog: "I'm pleased to note I called this one last December, when Anthropic first declared skills to be an "open standard"" (simonwillison.net).

For reactions on X, we rely on daily.dev's roundup (daily.dev). On the welcoming side, Blum_OG wrote: "Donating the standard to a foundation is a strong move for keeping it open and neutral." The cautious voices were specific. kbhaskar_95 said skill distribution should borrow what package registries learned the hard way: "signed publishers, pinned versions, a diff on update." BitsandMettle wrote that "a version field in the SKILL.md frontmatter would help a lot once more agents load the same files." illulachy_ called it an open question whether agents "behave the same on one skill file." The parse failure above is a small example of exactly that.

What this means for developers in Korea

Nothing changes right away. Existing skills keep working, and you can keep using and contributing to the standard without joining AAIF.

Over time, two things look likely to shift. First, a skill written once is more likely to work beyond Claude Code, in tools like ChatGPT, Cursor, and GitHub Copilot. Korean teams packaging internal procedures as skills have less reason to worry about lock-in. Second, once the spec gains versions and security guidance, files that rely on lenient parsing may get filtered out. If a missing pair of quotes is keeping a file from parsing, fixing it now is cheaper.

There is also a direct way to take part. Anthropic says feedback on proposed spec changes goes through the project's Discord. Guidance for writing descriptions in non-English languages seems like a natural proposal for Korean teams to bring.

For an earlier look at a single skill file on this site, see Claude Code TDD Skill review (in Korean).

Limitations

Sources

• Anthropic, Donating Agent Skills to the Agentic AI Foundation (2026-10-22): https://www.anthropic.com/news/donating-agent-skills-to-the-agentic-ai-foundation
• Anthropic, Donating the Model Context Protocol and establishing the Agentic AI Foundation (2025-12-09): https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation
• AAIF, Agentic AI Foundation Announces Four New Projects (2026-10-22): https://aaif.io/news/agentic-ai-foundation-announces-four-new-projects
• Agent Skills specification: https://agentskills.io/specification
• skills-ref reference implementation (parser.py): https://github.com/agentskills/agentskills/blob/main/skills-ref/src/skills_ref/parser.py
• Vercel, State of agent skills (2026-09-25): https://vercel.com/blog/state-of-agent-skills
• The Agent Forge, What 9,930 public agent skills tell your agent to do (2026-10-19): https://theagentforge.io/reports/skill-scan-2026-10
• Hacker News submission: https://news.ycombinator.com/item?id=50198911
• Simon Willison: https://simonwillison.net/2026/Oct/22/donating-agent-skills/
• daily.dev roundup: https://daily.dev/posts/openai-and-anthropic-donate-open-responses-and-agent-skills-to-the-agentic-ai-foundation-g0bsrnlul