system-create-skill
用途: 创建一个完整的 System/Project
触发: 当用户表达想要创建一个系统时(如:"创建一个健康管理系统")
核心理念:
- •人只说意图,AI 自己设计
- •做选择题,不做填空题
- •问最关键的一题(最大信息增益)
Workflow 定义
这是一个完整的 Workflow,演示了:
- •条件分支
- •循环迭代
- •Workflow 嵌套
- •判别式交互
yaml
workflow:
name: system-create-skill
description: 创建一个完整的 System/Project
version: 1.0.0
input:
user_intent: string # 用户的模糊意图
output:
system_path: string # 创建的系统路径
summary: object # 系统摘要
# ========================================
# 节点定义
# ========================================
nodes:
# ────────────────────────────────────
# 阶段 1: 快速澄清(最大信息增益)
# ────────────────────────────────────
clarify_intent:
type: skill
skill: clarify-skill
input:
question: $workflow.input.user_intent
mode: "multiple_choice"
strategy: "max_information_gain"
output_to: $clarified_requirements
# ────────────────────────────────────
# 阶段 2: 自主设计
# ────────────────────────────────────
design_system:
type: workflow
workflow: system-design-workflow
input:
requirements: $clarified_requirements
output_to: $design_plan
# ────────────────────────────────────
# 阶段 3: 展示方案(判别式确认)
# ────────────────────────────────────
show_plan:
type: skill
skill: user-confirm-skill
input:
content: $design_plan
format: "visual_summary"
question: "这个方案符合你的预期吗?需要调整哪里?"
output_to: $user_feedback
# ────────────────────────────────────
# 条件判断:是否通过
# ────────────────────────────────────
check_approval:
type: condition
expression: "$user_feedback.approved == true"
# ────────────────────────────────────
# 分支 A: 执行创建
# ────────────────────────────────────
execute_creation:
type: workflow
workflow: system-execute-workflow
input:
plan: $design_plan
output_to: $execution_result
# ────────────────────────────────────
# 分支 B: 调整设计
# ────────────────────────────────────
adjust_design:
type: skill
skill: iterate-skill
input:
artifact: $design_plan
feedback: $user_feedback.comments
output_to: $adjusted_plan
# ────────────────────────────────────
# 循环控制器
# ────────────────────────────────────
iteration_controller:
type: loop
max_iterations: 3
condition: "$user_feedback.approved == false"
# ────────────────────────────────────
# 阶段 5: 验证结果
# ────────────────────────────────────
validate_result:
type: skill
skill: user-confirm-skill
input:
content: $execution_result
format: "file_list_with_summary"
question: "试用一下,有没有不符合预期的地方?"
output_to: $validation_feedback
# ────────────────────────────────────
# 最终检查
# ────────────────────────────────────
final_check:
type: condition
expression: "$validation_feedback.satisfied == true"
# ────────────────────────────────────
# 如果不满意:记录反馈
# ────────────────────────────────────
log_feedback:
type: skill
skill: log-skill
input:
message: "System created but user requested improvements"
data: $validation_feedback
# ========================================
# 边定义(控制流)
# ========================================
edges:
# 主流程
- from: clarify_intent
to: design_system
- from: design_system
to: show_plan
- from: show_plan
to: check_approval
# 条件分支
- from: check_approval
to: execute_creation
condition: true
- from: check_approval
to: adjust_design
condition: false
# 循环:调整后重新进入设计
- from: adjust_design
to: iteration_controller
- from: iteration_controller
to: design_system
condition: "$loop.should_continue"
- from: iteration_controller
to: execute_creation
condition: "$loop.should_exit" # 超过最大次数,强制执行
# 验证流程
- from: execute_creation
to: validate_result
- from: validate_result
to: final_check
- from: final_check
to: END
condition: true
- from: final_check
to: log_feedback
condition: false
- from: log_feedback
to: END
# ========================================
# 入口和出口
# ========================================
entry: clarify_intent
exit: END
子 Workflow: system-design-workflow
yaml
workflow:
name: system-design-workflow
description: 设计系统的详细方案
input:
requirements: object # 澄清后的需求
output:
design_plan: object # 设计方案
nodes:
# 调研最佳实践
research:
type: skill
skill: research-skill
input:
topic: $workflow.input.requirements.domain
focus: "best_practices, patterns, tools"
# 制定计划
plan:
type: skill
skill: plan-skill
input:
goal: $workflow.input.requirements
knowledge: $research.output
# 设计目录结构
design_structure:
type: skill
skill: structure-design-skill
input:
plan: $plan.output
# 设计 Workflows
design_workflows:
type: skill
skill: workflow-define-skill
input:
requirements: $workflow.input.requirements
structure: $design_structure.output
# 设计 Skills
design_skills:
type: skill
skill: domain-skill-design
input:
workflows: $design_workflows.output
structure: $design_structure.output
# 设计 Agents(如果需要)
design_agents:
type: skill
skill: agent-define-skill
input:
skills: $design_skills.output
workflows: $design_workflows.output
# 汇总方案
synthesize:
type: skill
skill: synthesize-skill
input:
structure: $design_structure.output
workflows: $design_workflows.output
skills: $design_skills.output
agents: $design_agents.output
edges:
- from: research
to: plan
- from: plan
to: design_structure
- from: design_structure
to: design_workflows
- from: design_workflows
to: design_skills
- from: design_skills
to: design_agents
- from: design_agents
to: synthesize
- from: synthesize
to: END
entry: research
exit: END
子 Workflow: system-execute-workflow
yaml
workflow:
name: system-execute-workflow
description: 执行创建系统的文件和代码
input:
plan: object # 设计方案
output:
result: object # 执行结果
nodes:
# 创建目录结构
create_structure:
type: skill
skill: create-file
input:
plan: $workflow.input.plan.structure
# 创建 Workflows
create_workflows:
type: skill
skill: create-file
input:
files: $workflow.input.plan.workflows
# 创建 Skills
create_skills:
type: skill
skill: create-skill
input:
specs: $workflow.input.plan.skills
# 创建 Agents
create_agents:
type: skill
skill: create-file
input:
files: $workflow.input.plan.agents
# 创建 README 和文档
create_docs:
type: skill
skill: create-file
input:
files: $workflow.input.plan.docs
# 汇总结果
summarize:
type: skill
skill: summarize-skill
input:
created_files:
- $create_structure.output
- $create_workflows.output
- $create_skills.output
- $create_agents.output
- $create_docs.output
edges:
- from: create_structure
to: create_workflows
- from: create_workflows
to: create_skills
- from: create_skills
to: create_agents
- from: create_agents
to: create_docs
- from: create_docs
to: summarize
- from: summarize
to: END
entry: create_structure
exit: END
执行示例
用户输入
code
"我想要一个健康管理系统"
执行流程
1. clarify_intent(澄清意图)
Claude 输出:
code
我理解你想要健康管理系统。为了给你最合适的方案,我需要了解: 你想要的健康管理系统,最核心的目标是什么? A. 预防疾病 - 定期体检报告分析 - 健康指标追踪 - 风险预警 B. 提升体能 - 运动计划和追踪 - 进度可视化 - 目标管理 C. 改善习惯 - 饮食记录和分析 - 睡眠质量追踪 - 行为养成提醒 D. 慢病管理 - 血糖/血压监测 - 用药提醒 - 医生沟通记录 你倾向哪个?或者需要组合?
用户回答: A
澄清结果:
json
{
"domain": "health_management",
"focus": "disease_prevention",
"key_features": [
"checkup_analysis",
"health_indicators_tracking",
"risk_assessment"
],
"interaction_frequency": "daily"
}
2. design_system(自主设计)
执行:
- •research-skill: 调研健康管理最佳实践
- •plan-skill: 制定实施计划
- •design_structure: 设计目录结构
- •design_workflows: 设计 3 个核心 Workflows
- •design_skills: 设计 8 个 Skills
- •design_agents: 设计 1 个 Agent
设计方案:
yaml
system_name: health-system
structure:
- .claude/
- agents/
- health-advisor.md
- skills/
- checkup-analysis/
- health-indicators/
- risk-assessment/
- daily-review/
- weekly-report/
- workflows/
- daily-check.yaml
- weekly-report.yaml
- checkup-analysis.yaml
- data/
- profile.json
- indicators/
- checkups/
- outputs/
- reports/
workflows:
- name: daily-check
trigger: tick (21:00)
steps: [collect_data, analyze, generate_report, notify]
- name: weekly-report
trigger: tick (Sunday 20:00)
steps: [aggregate_week, analyze_trends, generate_insights, notify]
- name: checkup-analysis
trigger: event (file_uploaded)
steps: [parse_report, compare_history, assess_risks, generate_recommendations]
skills:
- checkup-analysis-skill
- health-indicators-skill
- risk-assessment-skill
- daily-review-skill
- weekly-report-skill
- data-collect-skill
- trend-analysis-skill
- notify-user-skill
agents:
- health-advisor (专长: checkup-analysis, risk-assessment, health-indicators)
3. show_plan(展示方案)
Claude 输出:
code
我已经完成设计!这是方案摘要: 📁 系统结构: - 1 个 Agent(健康顾问) - 8 个 Skills(体检分析、指标追踪等) - 3 个 Workflows(每日检查、每周报告、体检分析) 🔄 核心流程: ① 每天 21:00 自动检查当日数据,生成健康提醒 ② 每周日汇总本周情况,分析趋势 ③ 上传体检报告时,自动分析并更新健康档案 📊 数据管理: - 个人档案:基础信息、病史 - 指标记录:按日期存储 - 报告输出:Markdown 格式 这个方案符合你的预期吗?需要调整哪里? [详细设计方案已保存到: .sop-engine/temp/design-plan.yaml]
用户回答: 看起来不错,可以开始创建
4. execute_creation(执行创建)
执行:
- •创建目录结构
- •创建 3 个 Workflow YAML 文件
- •创建 8 个 Skill(每个包含 SKILL.md + criteria.md)
- •创建 1 个 Agent
- •创建 README.md 和 DESIGN.md
结果:
code
✅ 系统创建完成! 📁 health-system/ ├── .claude/ │ ├── agents/health-advisor.md │ ├── skills/ (8 个) │ └── workflows/ (3 个) ├── data/ │ ├── profile.json (示例) │ └── indicators/ ├── outputs/ └── README.md 🚀 下一步: 1. 填写 data/profile.json(你的基本信息) 2. 试用: "执行每日健康检查" 3. 上传体检报告测试分析功能
5. validate_result(验证)
Claude:
code
试用一下,有没有不符合预期的地方? 如果需要调整,告诉我具体哪里需要改进。
需要的新 Skills
为了让这个 Workflow 工作,我们需要创建以下新 Skills:
高优先级
- •user-confirm-skill - 判别式确认
- •structure-design-skill - 设计目录结构
- •domain-skill-design - 设计领域 Skills
- •synthesize-skill - 汇总信息
- •summarize-skill - 总结结果
中优先级
- •workflow-define-skill - 定义 Workflow(已在 TODO)
- •agent-define-skill - 定义 Agent(已在 TODO)
评价标准
见 criteria.md
使用方式
触发
code
用户: "创建一个健康管理系统" Claude: [自动触发 system-create-skill]
预期时长
- •快速澄清: 1-2 轮对话
- •自主设计: 自动执行(2-3 分钟)
- •展示确认: 1 轮对话
- •执行创建: 自动执行(1-2 分钟)
- •验证: 1 轮对话
总计: 约 5-10 分钟,3-5 轮对话
实现策略
使用 loop-skill
由于我们遵循 "skill 和文件是一等公民" 原则,实际执行时:
- •system-create-skill 本身是一个 Skill
- •它内部调用 loop-skill 来实现循环
- •使用文件存储状态(.sop-engine/running/)
状态文件
code
.sop-engine/running/system-create-20260119-001/
├── state.json # 当前阶段
├── context.json # 所有变量
└── logs/
├── clarify.log
├── design.log
└── execute.log
版本历史
- •v1.0.0 (2026-01-19): 初始版本