10 月 22 日，Anthropic 宣布把 Agent Skills 捐给 Agentic AI Foundation（AAIF）。AAIF 是 Linux 基金会之下的一个基金，去年 12 月由 Anthropic、Block、OpenAI 共同发起，此前已经托管了 MCP 和 AGENTS.md。按 Anthropic 给出的数字，Agent Skills 已被 70 多个 agent 平台接入，累计发布约 100 万个 skill，安装近 2.8 亿次。
本文先说明这次捐赠交出了什么，再整理规范本身的规定，然后用一个按规范逐条写的检查脚本，查了我本机 Claude Code 里的 65 个 skill，最后落到写 skill 的人需要调整什么。
这次捐赠在功能上没有带来任何新东西，规范文本、文件格式、存放位置都没变。变的是归属：skill 的格式从一家公司的产品约定，变成了 Google、OpenAI 等公司的工程师共同维护的标准。对写 skill 的人来说，这意味着 SKILL.md 里 name 和 description 两个字段的写法，从此是跨平台的契约。本地自查的结果说明，这份契约最容易被忽视的不是格式错误，而是体量：65 个 skill 里只有 3 个格式不合规，却有 13 个正文超过了规范建议的 5000 token。
Anthropic 交出的三样东西：名称、域名和 GitHub 组织
原文列得很具体，捐出的是三样东西：Agent Skills 的名称和 logo、GitHub 组织 agentskills，以及 agentskills.io 域名。规范文本继续放在 agentskills.io，地址不变。
维护方式随之改变。原文写明，标准今后由 Anthropic、Google、OpenAI 等公司的工程师共同维护，两位 Lead Maintainer 是 David Soria Parra 和 Mahesh Murag。路线图上有三件事：版本化发布、正式的提案流程、安全指南。原文还特意说明，开发者不需要加入 AAIF，也可以继续使用和参与贡献这个标准。
采用数据方面，原文点名的平台有 ChatGPT、Cursor、GitHub Copilot 和 VS Code。规范仓库 agentskills/agentskills 创建于 2025 年 12 月，采用 Apache-2.0 许可，截至 10 月 24 日有 26529 个星。
AAIF 托管的 4 个项目分别管 agent 的哪一层
AAIF 成立时有三个创始项目：Anthropic 的 MCP、Block 的 goose、OpenAI 的 AGENTS.md。加上这次的 Agent Skills，一共 4 个。
这 4 个项目分管 agent 的不同层面，彼此不重叠。MCP 规定 agent 怎么连接外部工具和数据；AGENTS.md 是放在仓库里的一份说明，告诉编程 agent 这个项目的规矩；Agent Skills 是按任务打包的能力，包括说明、脚本和参考资料；goose 是一个开源的 agent 框架本身。
前三个都是「agent 读什么」的格式约定，正是最需要跨平台统一的部分。一个团队如果同时用 Claude Code、Codex 和 Cursor，工具接入、仓库规矩、可复用的能力包这三样东西，现在都有了不归属于任何一家厂商的格式。
MCP 加入 AAIF 时的经验可以参考。MCP 官方博客当时写道，加入后「For MCP little changes」，决策仍然由原来的维护者做，社区通过 SEP 提案流程参与。Agent Skills 的路线图里也写了「正式提案流程」，可以对照 MCP 的 SEP 流程来理解。
规范的核心是三层渐进加载
规范对目录结构的要求很少：一个 skill 是一个目录，里面至少有一个 SKILL.md；scripts/、references/、assets/ 三个子目录是可选的约定。
SKILL.md 开头必须是 YAML frontmatter，必填字段只有两个。name 只能用小写字母、数字和连字符，长度 1 到 64，不能以连字符开头或结尾，不能出现连续两个连字符，并且必须和所在目录同名。description 长度 1 到 1024 字符，要写清楚这个 skill 做什么、什么时候该用。可选字段有 license、compatibility、metadata，以及标注为实验性的 allowed-tools。
规范真正约束的是加载方式，分三层：
- 启动时：所有 skill 的 name 和 description 都会被加载，规范给的量级是每个约 100 token。
- 激活时：agent 判断要用某个 skill，才读入 SKILL.md 全文，建议控制在 5000 token 以内、500 行以内。
- 需要时：scripts、references、assets 里的文件，只在任务用到时才读。
这套设计决定了 description 的分量。它是唯一一段对所有 skill 常驻上下文的文字，agent 只能靠它判断要不要激活这个 skill。写得太短，agent 匹配不到任务；写得太长，每多装一个 skill，所有对话都要为它多付一份 token。
本地 65 个 skill 的自查结果
规范仓库提供了官方校验工具 skills-ref。这次我没有运行官方工具，而是照着规范原文，把上面每一条约束写成了一个检查脚本，对本机 Claude Code 目录下的 65 个 skill 逐个检查。token 用 GPT-4o 的分词器估算，和 Claude 实际的分词数会有出入，只看量级。
格式层面的问题很少。65 个里有 62 个完全合规，3 个有问题：2 个目录里没有 SKILL.md，1 个 name 和目录名不一致。没有出现大写字母、下划线或超长 description。
问题集中在体量上：
- 13 个 skill 正文超过 5000 token，其中 7 个超过 500 行。它们一旦被激活，就会一次性占掉相当一块上下文。
- 启动时常驻的元数据合计约 7500 token。每个 skill 的 name 加 description 中位数 96 token，和规范给的约 100 token 吻合；最长的一个有 545 token。
- 6 种规范外的 frontmatter 字段，最多的是 origin，出现在 5 个 skill 里。规范给额外信息留的位置是 metadata 字段，直接写在顶层的字段，规范没有定义，换一个平台不一定能被识别。
另有 44 个 skill 带了 scripts、references 或 assets 子目录，把细节拆到子目录、正文只留步骤的写法已经很普遍。超长的那 13 个，正是规范建议把参考资料拆进 references/ 的对象。
对写 skill 的人有哪些变化
格式本身没有变化，已经写好的 skill 不需要迁移。值得调整的是三件事。
第一，把 description 当作跨平台的接口来写。规范已经由多家公司共同维护，name 和 description 的规则会是各平台都遵守的最小公约数。description 里写清楚「做什么」和「什么时候用」，比堆砌功能列表更容易被不同的 agent 正确匹配。
第二，把平台专属信息收进 metadata。像 origin 这类自定义字段，放在顶层只有当前平台认识；metadata 是规范里明确定义的扩展位置。
第三，把来路不明的 skill 当代码看待。skill 可以带可执行脚本，allowed-tools 字段还能预先批准工具调用。路线图把「安全指南」列为待办，说明这部分规范目前还没有定。在它发布之前，第三方 skill 装进来之前，至少读一遍 scripts 目录里的内容。
规范的硬性规则只有十来条，照着原文写一个检查脚本并不费事。给自己的 skill 目录做一次体检，先看格式错误，再看哪些正文超过 5000 token、该把参考资料拆进 references/。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
On 22 October, Anthropic announced that it would donate Agent Skills to the AAIF. AAIF is a fund under the Linux Foundation, co-sponsored by Anthropic, Block, Open AI last December, which previously hosted MCP and AGENTS.md. According to Anthropic, Agent Skills has been connected to more than 70 agent platforms, cumulatively publish has about 1 million skills and has been installed nearly 280 million times.
This is the first time that this donation has given up something, then it's the norm itself, and then a rule-by-script examination of 65 skills in Claude Code, and finally what needs to be adjusted by the person who wrote it.
The donation did not bring any new features functionally, nor did it change the normative text, the format of the document, the location of the repository. Changed to belonger: the skill format changed from a product contract of a company to a standard that was jointly maintained by engineers from Google, Open AI and others. For those who write Skill, this means that the SKILL.md name and description of two fieldes are written from now on as a cross-platform contract. The results of the local self-check show that the compact is most easily ignored not by misformation, but by volume: only 3 out of 65 skills are in non-compliance, while 13 texts exceed the normative recommendation of 5,000 token.
Anthropic handed over three items: name, domain name and GitHub
The original text is very specific, and three things are donated: the name and logo of Agent Skills, the GitHub organization angentskills, and andentskills.io domain names. The text of the norm continues to be available at the following address.
The way of maintenance has changed. The original text states that the standards will be jointly maintained by engineers from Anthropic, Google, Open AI, and that the two Lead Mainainer are David Soria Parra and Mahesh Murag. There are three things on the road map: a version of publish, a formal proposal process, a security guide. It is also specified that developers do not need to join AAIF and can continue to use and participate in the contribution criterion.
With regard to the use of data, original named platforms are ChatGPT, Cursor, GitHub Copilot and VS Code. Regulates the warehouse, which was created in December 2025, with Apache-2.0 permits, with 26529 stars as of 24 October.
AAIF hosting 4 projects that manage the level of angent
AAIF was founded with three founding projects: MCP of Anthropic, goose of Block, AGENTS.md of Open AI. And this time, Agent Skills, four.
These four projects cover the different dimensions of angent and do not overlap. MCP sets out how angent connects external tools and data; AGENTS.md is a note in the warehouse that informs the programming protocol of the project; Agent Skills is the task-driven ability to package, including instructions, scripts and references; goose is an open-source angent framework itself.
The first three were "agent reading what" arrangements, which are the parts that need to be integrated across the platform. A team that uses both Claude Code, Codex and Cursor, tool access, warehouse rules, reusable capacity packages now has formats that do not belong to any manufacturer.
The experience of MCP in joining AAIF may be consulted. The MCP official blog then wrote that after the addition of "For MCP little changees", decision-making is still done by the original defender and the community is involved through the SEP proposal process. Agent Skills' road map also contains a "official proposal process" that can be understood against the MCP's SEP process.
The core of the norm is three layers of incremental loading.
The normative requirements for the catalogue structure are minimal: one skill is a catalogue containing at least one SKILL.md; scripts/, references/, asets/ subdirectories are optional engagements.
SKILL.md must start with YAML Frontmatter, only two fields. Name can only be used in lowercase letters, numbers and hyphens, with 1 to 64 lengths, and cannot start or end with hyphens, without two consecutive hyphens, and must have the same name as the directory. Decryption 1 to 1024 characters, write clearly what this skill does and when to use. field is available for license, compatibility, metadata, as well as allowed-tools marked as experimental.
The regulation is based on a three-tiered approach:
- On startup: All skill 's name and description will be loaded, and the standard is about 100 token each.
- On activation: angent judge to use a skill to read the full SKILL.md, recommended to be contained within 5,000 token and within 500 rows.
- When necessary: documents in scripts, references, asets, read only when the task is used.
This design determines the weight of the design. It's the only piece of text that all skills permanent to context can only rely on to determine whether to activate this skill. It's too short, angent doesn't match the task; it's too long, it's too long, and every more skill, all the conversations pay it an extra token.
Self-check of 65 skills locally
Standardized warehouses provide official verification tools, skills-ref. I didn't run the official tools this time, but I wrote each of the above constraints into a check script, one by one, of 65 skills under my own Claude Code directory. Token uses GPT-4o's symmetry, and Claude's actual symmetrical number will differ.
There are few problems at the format level. Sixty-two out of 65 are fully compliant and three are problematic: two directories do not have SKILL.md, one name does not match the directory name. No capital letters, underlineds or excessive length appear.
The problem is in volume:
- 13 skill text over 5,000 token, of which 7 over 500 rows. Once they're activated, they take a piece of context once and for all.
- A total of approximately 7,500 tokens of metadata that are permanently held at start-up. The name of each skill plus the median number of 96 token, which corresponds to approximately 100 token; the longest is 545 token.
- 6 Frontmatter field outside the norms, most likely origin, in 5 skills. Regulation of the location of additional information is metadata field, written directly at the top level of field, the norm is not defined and the replacement platform may not be identified.
A further 44 skills have a subdirectories of scripts, references or asets, and it is common to remove the details to the subdirectories, where only steps are taken. The very long 13 are the subject of a normative proposal to tear references down.
What's the difference with the person who wrote Skill?
The format itself has not changed and the already written skill does not need to move. There are three things worth adjusting.
First, write as a cross-platform interface. The regulation has been jointly maintained by a number of companies, and the rules of name and description would be the smallest number of conventions to be observed by each platform. Discription clearly states "do what" and "when" and is easier to match than stacking functional lists.
Second, the Platform's exclusive information is integrated into the metadata. Self-defined fields such as origin, which are only known on the top of the current platform; metadata is a clearly defined extension of the norm.
Third, put the unknown skill when the code looks. Skill can bring enforceable scripts and allow advance approval of tool call. The road map lists the "security guide" as pending, indicating that this part of the norm is still pending. Before it's publish, the third party skill at least read what's in the scripts directory.
There are only a dozen mandatory rules, and it is not difficult to write a script in the original language. Take a medical examination of your skill directory, first looking at formatting errors, then looking at which text exceeds 5000 token and which references should be removed into references/.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.