Agentic Skills Framework：Agent技能框架方法论

什么是Agentic Skills Framework

Agentic Skills Framework将Agent能力组织为四个层级：

核心原则

• 组合优于继承：Skills通过组合形成更强大的能力
• 显式优于隐式：每个Skill的输入输出明确声明
• 测试即文档：Skills自带测试用例作为使用文档
• 失败即学习：Skill执行失败自动触发改进流程

Skill解剖学

一个完整的Skill包含以下结构：

# superpowers skill结构
skills/
└── web-search/
    ├── SKILL.md           # 技能描述文档
    ├── schema.json        # 输入输出Schema
    ├── handler.py         # 执行逻辑
    ├── tests/             # 测试用例
    │   ├── basic.test.py
    │   └── edge.test.py
    ├── examples/          # 使用示例
    │   └── example1.yaml
    └── dependencies/      # 依赖的子Skills
        └── http-request/

Schema定义

# schema.json - Skill接口定义
{
  "name": "web-search",
  "version": "1.2.0",
  "description": "搜索互联网并返回结构化结果",
  "inputs": {
    "query": {
      "type": "string",
      "required": true,
      "description": "搜索关键词"
    },
    "max_results": {
      "type": "integer",
      "default": 10,
      "description": "最大结果数"
    },
    "sources": {
      "type": "array",
      "items": { "type": "string" },
      "default": ["google", "bing"],
      "description": "搜索引擎列表"
    }
  },
  "outputs": {
    "results": {
      "type": "array",
      "items": {
        "title": "string",
        "url": "string",
        "snippet": "string",
        "relevance_score": "number"
      }
    },
    "metadata": {
      "total_found": "integer",
      "search_time_ms": "integer",
      "sources_used": "array"
    }
  },
  "dependencies": ["http-request", "html-parser"],
  "cost": {
    "api_calls": 2,
    "estimated_time_ms": 1500
  }
}

Skill组合模式

1. 串行组合（Pipeline）

# 串行执行：搜索 → 提取 → 总结
pipeline:
  - skill: web-search
    args: { query: "${user_query}" }
    output_var: search_results
  
  - skill: content-extractor
    args: { urls: "${search_results.urls}" }
    output_var: content
  
  - skill: summarizer
    args: { text: "${content}" }
    output_var: summary

2. 并行组合（Parallel）

# 并行执行：同时搜索多个来源
parallel:
  - skill: web-search
    args: { query: "${query}", sources: ["google"] }
    output_var: google_results
  
  - skill: web-search
    args: { query: "${query}", sources: ["bing"] }
    output_var: bing_results
  
  - skill: web-search
    args: { query: "${query}", sources: ["duckduckgo"] }
    output_var: ddg_results

merge:
  skill: result-merger
  args: { results: ["${google_results}", "${bing_results}", "${ddg_results}"] }

3. 条件组合（Conditional）

# 条件执行：根据错误类型选择策略
conditional:
  skill: error-handler
  branches:
    - condition: "error.type == 'timeout'"
      skill: retry-with-backoff
      args: { max_retries: 3, delay_ms: 1000 }
    
    - condition: "error.type == 'rate_limit'"
      skill: rate-limit-handler
      args: { wait_minutes: 5 }
    
    - condition: "error.type == 'auth_failure'"
      skill: credential-refresher
      args: { credential_store: "vault" }
    
    - default:
      skill: escalate-to-human

4. 循环组合（Loop）

# 迭代优化：多轮改进
loop:
  skill: iterative-improver
  max_iterations: 5
  termination:
    - quality_score >= 0.9
    - iteration_count >= 5
  
  body:
    - skill: evaluator
      args: { content: "${current_draft}" }
      output_var: evaluation
    
    - skill: improver
      args: { 
        content: "${current_draft}", 
        feedback: "${evaluation.feedback}" 
      }
      output_var: improved_draft

OpenClaw集成方案

1. 转换为OpenClaw Skill格式

# ~/.openclaw/skills/web-search/SKILL.md
---
name: Web Search Pro
description: 增强版网页搜索，支持多引擎并行
version: 1.2.0
trigger: search the web
cost: ~$0.002/call
---

## 输入参数
- query (string, 必需) - 搜索关键词
- max_results (number, 默认10) - 最大结果数

## 输出格式
返回结构化搜索结果列表，每条包含：
- title: 标题
- url: 链接
- snippet: 摘要
- relevance: 相关度评分

## 依赖Skills
- http-request: HTTP请求工具
- html-parser: HTML解析器
- rate-limiter: 限流器

## 最佳实践
- 使用具体关键词，避免模糊查询
- 并行调用多个搜索引擎提高覆盖率
- 结果缓存24小时避免重复调用

2. 使用Superpowers方法论

# OpenClaw项目中的Skill组织
~/.openclaw/skills/
├── primitives/           # L1: 原子工具
│   ├── http-request/
│   ├── file-read/
│   ├── shell-exec/
│   └── database-query/
├── compound/             # L2: 复合技能
│   ├── web-search/       # 组合: http-request + html-parser
│   ├── code-review/      # 组合: file-read + analyzer
│   ├── deploy/           # 组合: shell-exec + http-request
│   └── monitor/          # 组合: database-query + alert
└── workflows/            # L3: 工作流
    ├── content-pipeline/ # 搜索→提取→生成→发布
    ├── dev-workflow/     # 编码→测试→审查→部署
    └── research-flow/    # 文献→总结→分析→报告

最佳实践

相关资源