On October 22 Anthropic announced it is donating the Agent Skills standard to the Agentic AI Foundation (AAIF), which sits under the Linux Foundation. Agent Skills is a packaging format: one folder with a SKILL.md, optionally with scripts and reference files, so an agent loads those capabilities only when it needs them.

Two things matter for you. First, the format did not change, so your existing skills stay as they are. Second, something that can carry scripts and works across several products is effectively a new software distribution channel, so "look before you install" is worth more now than it was.

Sources come in two tiers. The Linux Foundation release and the agentskills.io spec are primary. Figures like "70+ platforms, one million skills, nearly 280 million installs" come from a third-party blog relaying Anthropic, and I could not confirm them on a primary page. They are flagged below.

A skill is a directory. The minimum is one file:

skill-name/
├── SKILL.md     # required: metadata + instructions
├── scripts/     # optional: executable code
├── references/  # optional: documentation
└── assets/      # optional: templates, resources

A concrete case: you write a "handle PDFs" skill. Day to day the agent only remembers its name and a one-line description. When you say "merge these PDFs", the agent reads the SKILL.md body and, if needed, runs a script from scripts/. When unused, it costs almost no context.

Everything below is from agentskills.io/specification. SKILL.md must start with YAML frontmatter, followed by Markdown content.

name (required)

Max 64 characters; lowercase letters, numbers and hyphens only; no leading or trailing hyphen, no consecutive hyphens; must match the parent directory name.

description (required)

1 to 1024 characters. Say both what the skill does and when to use it. The spec’s good example has specific keywords; its poor example is "Helps with PDFs."

license, compatibility, metadata (optional)

compatibility is up to 500 characters for environment requirements; metadata maps string keys to string values, and the spec suggests fairly unique key names to avoid clashes.

allowed-tools (optional, experimental)

A space-separated list of pre-approved tools, such as Bash(git:*) Bash(jq:*) Read. The spec says support may vary between agents.

The body has no format restrictions. The spec recommends step-by-step instructions, input/output examples and edge cases; keep SKILL.md under 500 lines and split longer material into referenced files, one level deep.

The spec splits loading into three levels, which is what Figure 1 shows. At startup each skill contributes only name and description, about 100 tokens. When the agent decides to use a skill it reads the full SKILL.md, and the spec advises keeping the body under 5000 tokens. Scripts, reference docs and assets load later, only when actually needed.

That is why description matters so much: the agent uses it to decide whether to pull the skill in. A vague description means a good skill never gets remembered.

The primary source is the Linux Foundation release (dated October 22, San Jose, timed with AGNTCon + MCPCon North America). It says Anthropic donated Agent Skills to AAIF, whose governance is described as open and vendor-neutral. Three more projects joined the same day: Open Responses from OpenAI (a spec for interoperable LLM interfaces), OpenEnv from Hugging Face (an RL environment framework) and OpenSandbox from Alibaba (a sandbox platform for AI apps). AAIF now hosts 10 projects, up from three at its December 2025 launch; existing ones include A2A, Agent Router, agentgateway, AGENTS.md, goose and MCP.

The following details come from third-party blogs (Developers Digest, OrcaRouter) and I did not see them in the release, so treat them as secondhand: the assets handed over are the name, logo, GitHub organization and agentskills.io domain; engineers from Anthropic, Google, OpenAI and others will maintain it as individuals; David Soria Parra and Mahesh Murag are lead maintainers; the roadmap proposes versioned releases, a formal proposal process and security guidance for skill authors and installers, with no dates and no version number yet.

Simon Willison notes he predicted this move in December 2025, when MCP had just joined AAIF. Anthropic’s own wording commits to keeping Agent Skills "open, community-driven, and vendor-neutral". Per Simon, the post gives no details on governance structure, the spec repo or how contributions work.

A third-party blog relaying Anthropic says more than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot and VS Code, have integrated Agent Skills since December, with one million published skills and nearly 280 million installs. These are vendor-reported figures; the blog says it did not verify them independently, and "integrated" is a count that says nothing about depth.

My stance: that the format works across several products is believable; the scale figures I would not quote until there is a primary page.

Because skills can carry scripts/ and can be copied between different vendors’ agents, one bad skill is a cross-vendor supply-chain risk. A third-party blog cites an independent scan: of 9,930 public SKILL.md files it found no malware but flagged 30 (0.33%), mostly two habits, piping downloads straight into a shell and reading whole key files. The scanner’s makers offer it for free and the blog notes it can be evaded, so a low flag rate says little about private or newly published skills.

Before installing a skill I do these things. They are common sense, not an official checklist:

• Read SKILL.md and every script in scripts/, looking for curl ... | sh and anything that reads ~/.ssh or .env.
• Check allowed-tools: what does it pre-approve? The spec calls the field experimental and agents handle it differently, so do not count on it being enforced uniformly.
• Run it first in a sandbox or a throwaway directory. Claude Code sandbox settings are in today’s Claude lesson.
• Trust only authors or repos you recognize, and pin to a commit instead of following main.

A minimal skill looks like this (the directory must be named pdf-merge, matching name):

pdf-merge/
└── SKILL.md

---
name: pdf-merge
description: Merges multiple PDF files into one. Use when the user asks to combine, join or concatenate PDFs.
---

1. List the PDF files the user named, in order.
2. Run scripts/merge.py with those files.
3. Report the output path.

Validate with skills-ref from the spec repo:

skills-ref validate ./pdf-merge

It checks the frontmatter and naming rules. A third-party blog advises reading the README before adding it to CI, since it did not run the command; neither did I. The skill above is adapted from the spec’s minimal example, and scripts/merge.py does not exist here; you would write it yourself.

• The body of Anthropic’s own announcement: the fetch failed, so I only have the headline and a one-line summary.
• AAIF’s formal governance documents and the spec repo’s contribution rules: the release does not say, and third-party blogs call the roadmap "proposed".
• How far OpenAI Codex supports Agent Skills: no primary statement found, so no conclusion.
• 70+ platforms, one million skills, 280 million installs: only secondhand.

This barely touches the format; it touches habits. A skill used to feel like one product’s feature. Now it is a public format with a foundation behind it, so write-once-use-anywhere is worth more and a mistake costs more. If you write skills: make description specific, move long material into references/, and keep scripts readable at a glance. If you install other people’s skills: treat it like installing software.

One more point: Anthropic still holds the lead maintainer seats (per a third-party blog), so "neutral" is partly formal for now. I think that depends on how the proposal process actually runs, and it is too early to call.

AAIF

Agentic AI Foundation, a Linux Foundation body hosting open agent projects such as MCP, AGENTS.md and Agent Skills.

frontmatter

YAML metadata at the top of a file, wrapped in ---.

progressive disclosure

Load the least information first and more only when needed.

Lead Maintainer

A project’s main maintainer, with a large say in direction.

• Agent Skills specification (agentskills.io)
• Agentic AI Foundation Announces Four New Projects (Linux Foundation, 2026-10-22)
• Donating Agent Skills to the Agentic AI Foundation (Simon Willison, 2026-10-22)
• Anthropic Newsroom (headline only)
• Agent Skills Moves to AAIF: What Changes (Developers Digest, secondhand)
• What Moved and What Did Not (OrcaRouter, secondhand)

Sources marked secondhand: verify against the originals.