技能组织者
- 作者仓库星标 1,039
- 作者仓库 MassGen
- 领域
- 人工智能
- 作者声明的智能体
- 未找到明确声明;不据此推断已兼容或已测试
- 静态检查
- 88 / 100 · 启发式扫描,不代表运行安全
- 运行环境证据
- 未执行;当前没有隔离运行测试证据
- 作者 / 版本 / 许可证
- @massgen · 未声明许可证
- 上下文用量评级
- 低消耗
- 接入复杂程度
- 即装即用
- 是否需要外部接口密钥
- 不需要
- 兼容的系统
- 未声明(默认跨平台)
- 运行环境要求
- 未检出明确要求
- 检测到的文件与系统行为
-
- 只读
- 允许写入 / 修改
- 网络行为
- 仅限本地
- 安装命令数
- 无(仅作为资料)
需要注意: 未声明工具权限范围;实际权限由运行环境决定。
第一步:库存
列出 .agent/skills/ 文件夹中的所有技能目录。然后阅读每个技能的 SKILL.md 以了解范围、质量和重叠。每个技能都位于 .agent/skills/ 的子目录中,并带有 SKILL.md 文件。
第 2 步:识别重叠或易混淆的技能
查找: 执行相同操作但名称略有不同的技能 范围明显重叠的技能(一个是另一个的子集) 可以组合成具有多个部分的单个更广泛技能的技能 从不同日志分析会话创建的近乎重复的技能
第 3 步:合并到分层父级技能中
对于每组重叠或相关的技能,创建一个包含涵盖每个子能力的部分的单父技能:选择更广泛的父名称和目录(例如,`web-app-dev`而不是单独的`react-frontend`,`nodejs-backend`,`web-testing`)编写一个全面的SKILL.md,并为每个技能明确标记部分
步骤4:生成SKILLREGISTRY.md
将`.agent/skills/SKILL_REGISTRY.md`编写为紧凑的路由指南:注册表应该: 按目的/领域对技能进行分组(不是按字母顺序) 对于每项技能,包括:它的用途、何时阅读以及它包含哪些部分 “何时使用”行至关重要 - 它告诉代理何时加载完整的 SKILL.md
第五步:报告
总结您所做的事情: 找到了多少技能 合并了哪些技能(旧名称 -> 新名称) 哪些技能保持原样 最终注册表结构
# Skill Organizer
Reads all installed skills, identifies overlaps and confusability, merges where
appropriate, and produces a compact `SKILL_REGISTRY.md` that serves as a routing
guide in agent system prompts.
## Why This Exists
Models have a skill selection capacity (kappa) of approximately 50-100 items.
Beyond that, selection accuracy degrades. As skills accumulate through log analysis
and manual creation, the skill library can cross this threshold. This skill
keeps the library organized and the routing guide compact.
Reference: [Scaling LLM Agents with Skill Libraries](https://arxiv.org/pdf/2601.04748v2)
## Workflow
### Step 1: Inventory
List all skill directories in the .agent/skills/ folder.
Then read each skill's SKILL.md to understand scope, quality, and overlap.
Each skill lives in a subdirectory of .agent/skills/ with a SKILL.md file.
### Step 2: Identify overlapping or confusable skills
Look for:
- Skills that do the same thing with slightly different names
- Skills whose scopes significantly overlap (one is a subset of another)
- Skills that could be combined into a single broader skill with multiple sections
- Near-duplicate skills created from different log analysis sessions
### Step 3: Merge into hierarchical parent skills
For each group of overlapping or related skills, create a single **parent skill**
with sections covering each sub-capability:
1. Choose a broader parent name and directory (e.g., `web-app-dev` instead of
separate `react-frontend`, `nodejs-backend`, `web-testing`)
2. Write one comprehensive SKILL.md with clearly labeled sections for each
sub-capability. Each section should be self-contained enough that an agent
can read just that section for a focused task.
3. Move bundled resources (templates, examples, configs) from merged skills
into subdirectories of the parent skill directory.
4. Remove the redundant skill directories.
When merging, prefer the skill with:
- Better-quality instructions and examples
- More complete bundled resources
- A more descriptive, general name
The goal is fewer, richer skills — a parent skill with 4 well-written sections
is better than 4 separate shallow skills.
### Step 4: Generate SKILL_REGISTRY.md
Write `.agent/skills/SKILL_REGISTRY.md` as a compact routing guide:
```markdown
# Skill Registry
## <Category> (<count>)
- **skill-name**: What it does in one sentence.
Use when: <trigger condition — when should the agent read this skill?>
Sections: <comma-separated list of sections/sub-capabilities within the skill>
- **skill-name**: What it does in one sentence.
Use when: <trigger condition>
## Recently Added
- **new-skill** (project): Frontmatter description — not yet categorized
```
The registry should:
- Group skills by purpose/domain (not alphabetically)
- For each skill, include: what it does, when to read it, and what sections it contains
- The "Use when" line is critical — it tells agents when to load the full SKILL.md
- The "Sections" line (when applicable) shows what sub-capabilities live inside the skill
- Stay under 50 entries total (merge aggressively if needed)
- Include a "Recently Added" section for skills created since last organization
- NOT duplicate full skill content, just enough for selection routing
### Step 5: Report
Summarize what you did:
- How many skills were found
- Which skills were merged (old names -> new name)
- Which skills were kept as-is
- The final registry structure
## Constraints
- Do NOT use keyword matching, Jaccard similarity, or heuristic categorization.
Use your understanding of what each skill does.
- Be aggressive about merging: fewer high-quality skills beats many overlapping ones.
- Preserve all bundled resources (templates, examples, configs) during merges.
- The SKILL_REGISTRY.md is a routing guide, not documentation. Keep it concise.
- Skills in `massgen/skills/` (builtin) should not be merged or deleted, only
cataloged in the registry.
章节与文件索引
按说明文件中的顺序排列。
为什么存在
第一步:库存
第 2 步:识别重叠或易混淆的技能
第 3 步:合并到分层父级技能中
步骤4:生成SKILLREGISTRY.md
文件/命令 -> SKILL_REGISTRY.md · web-app-dev · react-frontend · nodejs-backend · web-testing · .agent/skills/SKILL_REGISTRY.md · massgen/skills/
内容校验值 -> 9f137502e2b4
当前原文的自动摘译,尚未人工审校。
为什么存在
模型的技能选择能力 (kappa) 约为 50-100 项。除此之外,选择准确性会降低。随着技能通过日志分析和手动创建不断积累,技能库可以跨越这个门槛。这项技能可以使图书馆井井有条,路线指南紧凑。参考:使用技能库扩展 LLM 代理