Agent Skills becomes a Linux Foundation standard, letting one SKILL.md run across agents

Anthropic hands the spec to a neutral home

On 22 October 2026, Anthropic donated the Agent Skills specification to the Agentic AI Foundation (AAIF), the Linux Foundation's home for agentic standards, according to a detailed write-up on dev.to. The donation covered the Agent Skills name, the file format, its GitHub organisation and the agentskills.io reference site. The handover was announced at AGNTCon + MCPCon North America in San Jose, and it was Anthropic's second contribution to the foundation after MCP.

AAIF took on several other projects the same day — OpenAI's Open Responses, Hugging Face's OpenEnv and Alibaba's OpenSandbox — growing from three hosted projects to ten. At the handover, dev.to reports more than 70 platforms integrated with the Agent Skills format, over one million skills published and roughly 280 million installs.

Who maintains it now

Stewardship sits with AAIF, but day-to-day maintenance is carried out by engineers from Anthropic, Google, OpenAI and other companies contributing as individuals; David Soria Parra and Mahesh Murag are named as lead maintainers. Anthropic has proposed a starting roadmap of versioned releases, a formal process for proposing changes, and published security guidance for people who write and install skills. The dev.to piece reads that last item as an official acknowledgement of a trust problem: at a million published skills and counting, install-time safety can no longer be tribal knowledge.

One thing that did not move is ownership. A skill you write remains your file, in your repository, under your licence — no foundation membership, paid account or permission is required to publish one.

What a skill is, in brief

An Agent Skill is a folder containing a SKILL.md file: YAML frontmatter, where name and description are required, plus plain-language instructions, optionally supported by scripts/, references/ and assets/ subfolders. The cost-saving mechanism is progressive disclosure. At startup an agent reads only each skill's name and description — a few dozen tokens — and loads the full instructions only when a request matches that description.

Two consequences follow, per the dev.to guide. First, the description is effectively the product: it decides whether a skill ever fires, so it should be written as a trigger rather than a summary, with keywords front-loaded. Some agents cap how much description text they ingest; Codex, for example, reportedly budgets roughly 2% of its context window for skill descriptions across all installed skills. Second, skills are cheap to keep around — dozens can sit installed for the price of a paragraph of context, unlike an always-loaded system prompt.

One spec, many folders

The specification is now shared, but the folders agents scan are not. Claude Code, Copilot, Codex, Cursor, Gemini CLI and OpenCode each read their own directories, so portability is a file-organisation problem rather than a format problem.

The recommended setup in the guide is a single canonical copy: keep the master skill in the shared .agents/skills folder, which the standard points at and several agents already read directly, then symlink or copy it into the per-agent folders — ~/.claude/skills for Claude Code, ~/.copilot/skills or ~/.agents/skills for Copilot, and so on. Codex, Cursor and Gemini CLI read the shared .agents/skills location directly, and OpenCode also falls back to it. On Windows, where symlinks require developer mode or admin rights, a directory junction or a small copy script run after edits works instead. Vendors are converging on .agents/skills, the guide notes, with older per-tool folders kept as compatibility aliases.

Writing a portable skill

The spec defines a small set of frontmatter fields every agent understands: name (1–64 characters, lowercase letters, numbers and single hyphens, matching the folder name) and description are required, while license, compatibility and metadata are optional but portable. The allowed-tools field is only partially honoured — some agents respect it, others ignore it — so it should never be relied on alone. Tool-specific frontmatter belongs out of the shared file entirely: some agents ignore it and some upload flows, including claude.ai's, reject it outright.

For the body, the guide suggests structuring instructions as when to use (and when not to), numbered steps, exact output format and edge cases, kept under 500 lines or roughly 5,000 tokens. Long reference material belongs in references/ files the agent opens only when needed, and scripts/ should be reserved for deterministic steps such as date maths or format conversion — every extra script is also attack surface.

Why it matters

Agent Skills stops being a Claude feature that other tools copied and becomes shared infrastructure governed by a neutral foundation, with backing from rival vendors. For developers and teams, the practical payoff is write-once reuse: a single SKILL.md can now serve Claude Code, Codex, Cursor, ChatGPT and Copilot, with the remaining friction reduced to symlinks and folder layout rather than format incompatibility. The coming versioning and security guidance matters just as much at this scale, since a million published skills make skill installation a supply-chain question, not a convenience feature.