Skill Evolver
Solve first. Materialize later.
Workflow
Phase 0: Setup Output Directory
Create a timestamped output directory for this session:
CODEBLOCK0
Tip: Use a short slug derived from the task (e.g., pdf-translate, data-export, api-integration)
Store the output path for subsequent phases:
CODEBLOCK1
Phase 1: Intent Analysis
Analyze the user task and output
${OUTPUT_DIR}/01-intent.md.
See template:
references/templates/01-intent.md
Phase 2: Skill Search
Follow the complete skill search workflow:
→ references/skill-search.md
This workflow covers:
- - CLI prerequisites and installation
- Local + Registry (dual-track) search
- Skill selection checkpoint
- Installation and verification
- Security audit
Output files:
- -
${OUTPUT_DIR}/02-candidates.md - Merged search results - INLINECODE5 - Installation verification (if installed)
- INLINECODE6 - Security audit report (if installed)
Phase 3: Deep Inspection
For each candidate skill, perform deep analysis:
Follow the workflow: references/skill-inspector.md
Output: INLINECODE7
Checkpoint: Approach Decision
After inspection, evaluate whether skills can solve the task:
LLM evaluates:
- - Do skill capabilities match task requirements?
- Is modification needed?
- Is fusion beneficial?
LLM Recommendation:
- - Orchestration (skills match well, no major modification)
- Fusion (skills partially match, combining creates new value)
- Native (no suitable skills found)
Options for user:
- - A: Orchestration (LLM recommended)
- B: Skill Fusion (enter coding mode)
- C: Use native abilities instead
- D: Re-analyze (return to Phase 3)
Phase 3.5: Skill Fusion (Conditional)
Only if approach is Fusion
Follow the complete skill fusion workflow:
→ references/skill-fusion.md
This workflow covers:
- - Fusion spec design
- Invoke skill-creator
- Audit fused skill
Output files:
- -
${OUTPUT_DIR}/03-fusion-spec.md - Fusion specification - INLINECODE9 - Security audit (if fusion)
Phase 4: Orchestration
Design execution plan and output
${OUTPUT_DIR}/04-orchestration.md.
See template:
references/templates/04-orchestration.md
Checkpoint: Plan Confirmation
Use AskUserQuestion tool (or similar tool to Human-in-the-Loop) to confirm plan:
- - A: Proceed with this plan
- B: Modify the plan
- C: Show alternatives
- D: Additional requirements (then revise)
Phase 5: Execution
Execute the plan. For each step:
- - Native: use your own reasoning
- Skill: invoke the skill with appropriate input
Checkpoint: Materialization Decision
Use AskUserQuestion tool (or similar tool to Human-in-the-Loop) to ask about preservation:
- - A: Yes, create a new skill (invoke
skill-creator) - B: No, this was one-time
- C: Save as draft for later review
- D: Additional requirements (then adjust scope)
Principles
CODEBLOCK2
技能进化器
先解决,后具现。
工作流程
阶段 0:设置输出目录
为本次会话创建一个带时间戳的输出目录:
bash
格式:output/MM-DD-<功能标识>/
示例:output/03-09-pdf-translate/
mkdir -p output/$(date +%m-%d)-<功能标识>
提示:使用从任务中提取的简短标识(例如:pdf-translate、data-export、api-integration)
为后续阶段存储输出路径:
OUTPUT_DIR=output/<已创建目录>
阶段 1:意图分析
分析用户任务并输出 ${OUTPUT_DIR}/01-intent.md。
参见模板:references/templates/01-intent.md
阶段 2:技能搜索
遵循完整的技能搜索工作流程:
→ references/skill-search.md
该工作流程涵盖:
- - CLI 前提条件与安装
- 本地 + 注册表(双轨)搜索
- 技能选择检查点
- 安装与验证
- 安全审计
输出文件:
- - ${OUTPUTDIR}/02-candidates.md - 合并后的搜索结果
- ${OUTPUTDIR}/02-verify.md - 安装验证(如已安装)
- ${OUTPUT_DIR}/02-audit.md - 安全审计报告(如已安装)
阶段 3:深度检查
对每个候选技能进行深度分析:
遵循工作流程:references/skill-inspector.md
输出:${OUTPUT_DIR}/03-inspection.md
检查点:方法决策
检查后,评估技能能否解决任务:
LLM 评估:
- - 技能能力是否匹配任务需求?
- 是否需要修改?
- 融合是否有利?
LLM 推荐:
- - 编排(技能匹配良好,无需重大修改)
- 融合(技能部分匹配,组合创造新价值)
- 原生(未找到合适技能)
用户选项:
- - A:编排(LLM 推荐)
- B:技能融合(进入编码模式)
- C:改用原生能力
- D:重新分析(返回阶段 3)
阶段 3.5:技能融合(条件性)
仅当方法为融合时执行
遵循完整的技能融合工作流程:
→ references/skill-fusion.md
该工作流程涵盖:
输出文件:
- - ${OUTPUTDIR}/03-fusion-spec.md - 融合规范
- ${OUTPUTDIR}/03-fusion-audit.md - 安全审计(如进行融合)
阶段 4:编排
设计执行计划并输出 ${OUTPUT_DIR}/04-orchestration.md。
参见模板:references/templates/04-orchestration.md
检查点:计划确认
使用 AskUserQuestion 工具(或类似的人机交互工具)确认计划:
- - A:按此计划执行
- B:修改计划
- C:展示备选方案
- D:额外需求(随后修订)
阶段 5:执行
执行计划。对于每一步:
- - 原生:使用自身推理能力
- 技能:使用适当输入调用技能
检查点:具现决策
使用 AskUserQuestion 工具(或类似的人机交互工具)询问是否保留:
- - A:是,创建新技能(调用 skill-creator)
- B:否,仅此一次
- C:保存为草稿,稍后审查
- D:额外需求(随后调整范围)
原则
优先级:原生 > 编排 > 临时 > 持久
- - 简单任务优先使用原生能力
- 现有技能可解决时优先使用编排
- 仅在验证后且证明有复用价值时才具现
- 始终提供选项 [D] 以便补充输入
- 用户提供新信息时重新优化