Agent Skills 成为基础设施：Anthropic 将其捐赠给 AAIF

Yesterday (October 22, 2026) Anthropic announced it is donating Agent Skills to the Agentic AI Foundation (AAIF), under the Linux Foundation. It is the same foundation that took the Model Context Protocol, and Anthropic co-founded it with Block and OpenAI. The numbers in the announcement: more than 70 agent platforms have integrated skills since the standard was published in December 2025, including ChatGPT, Cursor, GitHub Copilot and VS Code. One million skills published, nearly 280 million installed. My read: companies tend to donate a format to a foundation when they want everyone else to stop worrying about who owns it. Experiments usually stay in-house. If that pattern holds here, this announcement is less about governance paperwork and more about Anthropic telling the industry to build on skills without hesitation. What a skill actually is A skill is a folder. That is the whole trick. pdf-processing/ ├── SKILL.md # required: metadata + instructions ├── scripts/ # optional: code the agent can run ├── references/ # optional: docs it can read on demand └── assets/ # optional: templates, data files Enter fullscreen mode Exit fullscreen mode SKILL.md is markdown with a small YAML block on top: --- name: pdf-processing description: "Extracts text and tables from PDF files, fills PDF forms, merges multiple PDFs. Use when working with PDF documents or when the user mentions PDFs, forms, or document extraction." --- ## Steps 1. ... Enter fullscreen mode Exit fullscreen mode Two fields are required: name and description. Everything else is optional. No runtime, no package manager, no registry account. If your agent can read files, it can in principle use a skill. The clever part is what does not get loaded The spec (agentskills.io) describes a three-level loading model: At startup the agent reads only the name and description of every skill, about 100 tokens each. When a task matches, it loads the full SKILL.md body (the spec recommends keeping it under 5000 tokens). Scripts, references and assets are pulled in only when the task actually needs them. This is the same idea as MCP tools, pushed one level up. Your agent can "know" hundreds of procedures without paying the context cost for any of them until the moment they matter. Context is the scarce resource in every agent loop, and skills are designed around that fact instead of against it. It also means the description field carries a lot of weight. It is the only part the agent sees before deciding to load anything. A vague description like "Helps with PDFs" gives the agent almost nothing to match a task against, so the skill may never trigger when it should. The good example in the spec reads like a search query the user might actually type. Why the donation matters MCP was donated to the same foundation earlier, and it is now supported across most major clients. To be fair, MCP's adoption was already well underway before the donation, so a foundation home is not a magic adoption switch. What it changes is the risk calculus: stewardship moves to a neutral body, with maintainers from Anthropic, Google, OpenAI and other companies acting as individuals. That makes it much harder for any single vendor to change the rules on a whim. Not impossible, foundations evolve too, but the incentives now point outward. Concretely, for anyone building with agents: A skill written to the spec should load in any client that integrated the standard: Claude Code, Cursor, Copilot, VS Code and dozens more. Support depth can vary between clients (the spec marks fields like allowed-tools as experimental), so test your skill where you actually run it. Companies can build products on the standard with less worry about one vendor's terms changing underneath them. My bet is the 70-platform number grows from here, because tools that were waiting to see who owned skills now have their answer. That is a prediction, not a promise. What I would do this week If you use an agent for anything repetitive: Pick one workflow you explain to your agent over and over. Writing PR descriptions, triaging issues, setting up a test env. Turn it into a folder with a SKILL.md. Steps, an example input and output, the edge cases you always forget. Spend most of your time on the description. Say what it does and when to use it, in words you would actually type. If the body grows past a page or two, move detail into references/ and let progressive disclosure do its job. The format took about a year to go from a Claude Code feature to a vendor-neutral standard with a million published skills. Writing good ones is still a rare skill. It probably will not stay rare for long. Written with AI assistance.