Skills

Definition

Skills are modular, reusable, evolvable units of agent capability. Unlike traditional tools (single-function APIs), Skills encode both what to do and how to behave — they include procedural instructions, domain knowledge, security constraints, and reference materials.

技能是代理能力的模块化、可复用、可演进单元。与传统工具（单功能 API）不同，技能不仅编码了**做什么**，还编码了**如何表现**——它们包含过程指令、领域知识、安全约束和参考资料。

Details

Skill vs. Related Concepts

Concept	What It Does	Analogy
Tool	Single function API (e.g., web search)	A hammer
MCP	Connection protocol between agent and tools	USB cable
Skill	Procedural knowledge + tools + constraints	A recipe
Harness	Runtime constraint and isolation layer	Safety goggles

| 概念 | 功能 | 类比 | |---------|-------------|---------| | **工具 (Tool)** | 单一功能 API（例如：网络搜索） | 锤子 | | **MCP** | 智能体与工具之间的连接协议 | USB 线 | | **技能 (Skill)** | 过程性知识 + 工具 + 约束 | 食谱 | | **约束框架 (Harness)** | 运行时约束与隔离层 | 护目镜 |

SKILL.md Structure (agentskills.io Standard)

A Skill is a folder containing a SKILL.md file with three layers:

1. Frontmatter — metadata: name, description, version, allowed-tools, compatibility
2. Body — procedural instructions the agent follows
3. References — supporting files loaded on demand

技能（Skill）是一个包含 `SKILL.md` 文件的文件夹，该文件由三个层级组成：1. **前言（Frontmatter）** — 元数据：名称、描述、版本、允许使用的工具（allowed-tools）、兼容性 2. **正文（Body）** — 智能体遵循的过程性指令 3. **参考资料（References）** — 按需加载的支持文件

Progressive Disclosure Principle

• ~100 tokens of metadata loaded at startup (always)
• <5000 tokens of instructions loaded on activation (on-demand)
• Reference files loaded only when needed (lazy)

- 启动时加载约 100 个词元的元数据（始终） - 激活时加载 <5000 个词元的指令（按需） - 引用文件仅在需要时加载（延迟）

Key Properties

• Portable: Write once, use across 20+ platforms (Claude Code, Cursor, VS Code, Gemini CLI, etc.)
• Composable: Skills can orchestrate other skills
• Evolvable: Frameworks like Memento-Skills enable skills to rewrite themselves from failures
• Securable: allowed-tools field provides native per-action permission whitelisting

- **可移植**：一次编写，即可在 20 多个平台（Claude Code、Cursor、VS Code、Gemini CLI 等）上通用 - **可组合**：技能能够编排其他技能 - **可演进**：借助 Memento-Skills 等框架，技能能够从失败中实现自我重写 - **可安保**：`allowed-tools` 字段提供了原生的单次操作权限白名单机制

Agent Capability Formula

Agent Capability = Base Model + Σ(Available Skills × Activation Probability)

智能体能力 = 基础模型 + Σ(可用技能 × 激活概率)

Connections

• Related to: Skill Lifecycle, Self-Evolving Agents
• Mentioned in: Skill Factory, agentskills.io, SkillNet

相关于：[[ai-agent-architecture/concepts/skill-lifecycle|技能生命周期]]，[[ai-agent-architecture/concepts/self-evolving-agents|自进化智能体]] - 提及于：[[ai-agent-architecture/sources/skill-factory-framework|技能工厂]]，[[ai-agent-architecture/sources/agentskills-io-ecosystem|agentskills.io]]，[[ai-agent-architecture/sources/skillnet-paper|SkillNet]]