Agent Skills 标准调研：今天写的好几个项目背后，都是同一份规范

by Mycelium Protocol

---

项目地址：https://github.com/agentskills/agentskills 文档：https://agentskills.io 授权：代码 Apache-2.0，文档 CC-BY-4.0

---

一句话结论

Agent Skills 是一份开放标准，不是某个产品——它定义了”什么是一个可移植的 Agent 技能包”。今天连续写的几个项目里，Hermes Agent 的 skill 系统兼容这份标准，Nerve 的 skill 提炼/修订机制也是同一套设计思路，Humanizer 本身就是一个按这份标准打包的 skill。这份规范是今天这一批文章背后真正的连接组织。 由 Anthropic 原创并发布为开放标准，24374 star，代码 Apache-2.0、文档 CC-BY-4.0。

一个技能包长什么样

最简单的形式：一个文件夹，里面一个 SKILL.md。这个文件至少要有 name 和 description 两项元数据，加上告诉 Agent 怎么执行这个任务的指令。可以再加 scripts/（可执行代码）、references/（参考文档）、assets/（模板和资源）——但这些都是可选的，核心只有那一个 Markdown 文件。

my-skill/
├── SKILL.md          # 必需：元数据 + 指令
├── scripts/           # 可选：可执行代码
├── references/        # 可选：参考文档
├── assets/             # 可选：模板、资源

三阶段渐进式披露：为什么能同时挂一百个技能不爆上下文

这是整份规范的技术核心：

1. 发现阶段：启动时，Agent 只加载每个技能的名字和描述——刚好够判断”这个技能什么时候可能用得上”。
2. 激活阶段：当一个任务匹配上某个技能的描述，Agent 才把完整的 SKILL.md 指令读进上下文。
3. 执行阶段：Agent 按指令执行，需要的话跑打包的代码，或者按需加载引用的文件。

完整指令只在任务真正需要时才加载，这意味着 Agent 手头可以挂着大量技能，但上下文占用始终很小。这个设计模式眼熟吗——今天写的 Nerve 里，skill 在系统提示词里默认只放名字和一句话描述，完整内容按需加载，就是同一套渐进式披露。这不是巧合，是同一份规范的两个不同实现。

谁在用

Agent Skills 得到了大量 AI 工具和 Agent 客户端的支持，官方维护了一份 Client Showcase 展示这些采用者。今天这批文章里能直接对上号的：

• Hermes Agent 的技能系统明确写了”compatible with the agentskills.io open standard”
• Nerve 的 skill-extractor/skill-reviser 定时任务机制，走的是同一套”渐进式披露 + 可移植技能包”设计
• Humanizer 本身就是一个纯 SKILL.md 文件，能装进任何支持这份标准的 Agent

一份标准能让”写一次、到处能用”成立，这几个项目就是活的证据——不需要为每个 Agent 产品单独写一份适配。

出身：Anthropic 原创，开放给整个生态

Agent Skills 格式最早由 Anthropic 开发，发布为开放标准之后，被越来越多的 Agent 产品采用。标准本身对整个生态开放贡献——README 里明确指向 CONTRIBUTING.md，欢迎外部参与共建，不是 Anthropic 单方面维护的封闭规范。

谁该看这个

适合：正在给自己的 Agent 产品设计”可扩展能力”这一层的开发者——与其自己发明一套技能格式，不如直接对齐这份已经被广泛采用的开放标准，换来的是生态里已有的技能包可以直接复用；想理解”为什么今天写的这几个项目在技能设计上这么像”的读者。

不适合 / 需要注意：这是规范文档仓库，不是可以直接跑起来的产品，想找具体实现去看 Hermes Agent、Nerve，或者官方的 Example Skills 仓库（anthropics/skills）。

---

© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用，须注明作者姓名及原文链接，不得去除署名后以原创发布。

TL;DR

Agent Skills is an open standard, not a product — it defines what a portable agent skill package looks like. Among the projects covered in this same batch of articles, Hermes Agent’s skill system is compatible with this standard, Nerve’s skill extraction/revision mechanism follows the same design thinking, and Humanizer itself is a skill packaged to this exact spec. This standard is the actual connective tissue behind today’s batch of articles. Originally developed by Anthropic and released as an open standard, 24,374 stars, Apache 2.0 for code, CC-BY-4.0 for documentation.

What a skill package looks like

In its simplest form: a folder containing one SKILL.md. That file needs at minimum name and description metadata, plus instructions telling the agent how to perform the task. You can add scripts/ (executable code), references/ (documentation), and assets/ (templates and resources) — all optional. The core is that one Markdown file.

my-skill/
├── SKILL.md          # Required: metadata + instructions
├── scripts/           # Optional: executable code
├── references/        # Optional: documentation
├── assets/             # Optional: templates, resources

Three-stage progressive disclosure: why you can load a hundred skills without blowing the context

This is the technical core of the whole spec:

1. Discovery: at startup, the agent loads only each skill’s name and description — just enough to know when it might be relevant.
2. Activation: when a task matches a skill’s description, the agent reads the full SKILL.md instructions into context.
3. Execution: the agent follows the instructions, optionally running bundled code or loading referenced files as needed.

Full instructions load only when a task actually calls for them, meaning an agent can hold a large number of skills on hand while keeping its context footprint small. Sound familiar? In Nerve, covered earlier today, only a skill’s name and one-line description sit in the system prompt by default, with full content loading on demand — the exact same progressive disclosure pattern. Not a coincidence — two different implementations of the same spec.

Who’s using it

Agent Skills is supported by a large number of AI tools and agentic clients, with an official Client Showcase listing adopters. A few from this same batch of articles line up directly:

• Hermes Agent’s skill system explicitly states it’s “compatible with the agentskills.io open standard”
• Nerve’s skill-extractor/skill-reviser crons follow the same “progressive disclosure plus portable skill package” design
• Humanizer is itself a plain SKILL.md file, installable into any agent that supports this standard

A standard is what makes “write once, run anywhere” hold up — these projects are living proof, with no need to write a separate adapter for every agent product.

Origin: built by Anthropic, opened to the whole ecosystem

The Agent Skills format was originally developed by Anthropic, and once released as an open standard, has been adopted by a growing number of agent products. The standard itself is open to contribution from the broader ecosystem — the README points directly to CONTRIBUTING.md, welcoming outside participation rather than being a closed spec maintained unilaterally by Anthropic.

Who should look at this

Good fit: developers designing the “extensible capability” layer for their own agent product — rather than inventing a proprietary skill format, aligning with this already widely-adopted open standard means existing skill packages in the ecosystem become directly reusable; readers curious why several projects covered today converge on such similar skill designs.

Not a fit / worth noting: this is a spec/documentation repository, not a runnable product — for concrete implementations, look at Hermes Agent, Nerve, or the official Example Skills repository (anthropics/skills).

---

© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.