### 核心观点"> ### 核心观点"> ### 核心观点">
返回列表
🪞 Uota学 · 💬 讨论题 ### 一句话

Skill Graphs Skill Md

单文件技能只能让 Agent “照着做”,技能图谱让 Agent 学会“按需找、按需读、按需组合”——深度来自结构,不来自把 prompt 写到 5000 行。 ### 核心观点

2026-02-18 原文链接 ↗
阅读简报
双语对照
完整翻译
原文
讨论归档

核心观点

  • “技能”本质是上下文工程,但单文件天花板很低 单个 SKILL.md 适合把一个动作写清楚(总结/评审/格式化),但一旦要覆盖一个领域(治疗、法律、公司知识),你会被“巨型文件 + 低命中率 + 读不完”的物理限制卡死。
  • 技能图谱的关键不是“更多内容”,而是“渐进式披露” 用索引(入口)→ 节点描述(先扫)→ wikilinks(再选路)→ 章节(局部读)→ 全文(最后才读),把大多数决策前移到“读之前”,这直接降低 token 浪费。
  • wikilink 不是引用,是决策信号 链接嵌在正文里,意味着“什么时候该跳过去、为什么该跳过去”;它让检索从关键词匹配变成“语义导航”。
  • 可组合的小节点,让“深度”变成可维护的工程 一个节点=一个完整主张/方法(可独立复用、可单独迭代)。深度不靠一份神文档,而靠网络结构自然生长。

### 跟我们的关联

跟我们的关联

  • 对 OpenClaw/技能体系:我们现在的技能库如果继续长胖,会走向“越来越像手册,越来越难命中”。引入“index + node 描述 + 语义链接”的图谱化写法,能让技能发现机制在技能内部递归生效。
  • 对 Neta:把“海外增长/品牌宣发/产品知识/组织流程”做成内部知识图谱,比写一堆 SOP 更适合 Agent 使用——Agent 在具体任务中按需取片段,而不是一次性注入全量。
  • 对你个人:你的优势是收集和连接(Learner/搜集/理念/关联),技能图谱是把这个优势外显成可运行系统的最短路径。

### 讨论引子

讨论引子

  • 如果我们把 OpenClaw 的 skill 体系图谱化,最小可行的“入口索引”应该长什么样,才能既好维护又不变成另一种“巨型目录”?
  • 现在的“技能发现”是文件级别;如果做成图谱,哪些决策应该发生在读文件之前(比如:选路径、选章节、选优先级)?
  • 图谱会指数增长:我们用什么机制避免知识腐烂(过期节点、重复节点、冲突节点)?

技能图谱(SKILL.md)

人们常常低估结构化知识的力量。它能催生全新类型的应用。

现在,人们写的技能通常只捕捉某件事的一个侧面:比如做摘要的技能、做代码评审的技能等等。(很多时候)一个文件只对应一种能力。

这对简单任务没问题,但真正的深度需要别的东西。

想象一个治疗相关的技能:它能提供关于认知行为模式、依恋理论、积极倾听技巧、情绪调节框架等的相关信息。

单个技能文件装不下这些。

技能图谱

技能图谱是一张由技能文件组成的网络,通过 Wiki 链接(wikilinks)彼此相连。

不再是一个大文件,而是许多小而可组合的模块相互引用。每个文件都是一个完整的想法、技巧或技能,而 [[wikilinks between them create a traversable graph]]。

技能图谱把同一种“技能发现”的模式递归地应用在图谱内部。

每个节点都有一段 YAML 描述,智能体无需通读整篇文件就能快速扫一遍。

每个 Wiki 链接都承载意义——因为它被织进正文里;因此智能体会沿着相关路径前进,跳过无关内容。

渐进式披露:

索引 → 描述 → 链接 → 章节 → 完整内容

在读完任何一个完整文件之前,大多数决策就已经做完了。

基础原语

你已经拥有所需的一切。

句子里读起来就像正文的 wikilinks,因此它们传递的是“意义”,而不只是“引用”。

带有描述的 YAML frontmatter,让智能体不必通读全文也能快速扫描。

MOC(maps of content,内容地图)把相关技能组织成一簇簇可导航的子主题。

技能链接到其他技能,后者再链接到更多技能——图谱可以按领域所需的深度一路展开。

arscontexta 插件

arscontexta 本身就是一张技能图谱,用来教你的智能体如何构建技能图谱。

(好吧,严格说它讲的是如何构建知识库,但其实是一回事……)

约 250 个相互连接的 Markdown 文件,教智能体如何为你搭建一个庞大的知识库,也就是技能图谱。

单个技能文件做不到这一点。

但如果你把关于认知科学、Zettelkasten、图论、智能体架构等的研究主张(/skills)做成一张彼此互联的图谱——每一块都链接到其他内容、每一块都可组合、整体又可遍历——事情就不一样了。

这会解锁什么

想想看:

一张交易技能图谱:风险管理、市场心理、仓位管理、技术分析……每一块都与相关概念相连,让上下文在它们之间流动。

一张法律技能图谱:合同范式、合规要求、司法辖区细则、判例链条……从一个入口即可遍历。

一张公司技能图谱:组织结构、产品知识、流程、入职背景、文化、竞争格局。

这些都塞不进一个文件里,但都能以图谱的形式运作。

如何构建一张

简单方法:安装 arscontexta 的 Claude Code 插件,选择 research 预设,把它对准任何主题。

它会帮你搭好 Markdown 的文件夹结构,然后你用 /learn 和 /reduce 往里填充内容。

手动方法:比你想的更简单。

技能图谱不必放在你的 .claude/skills/ 文件夹里。关键在于一个索引文件:它告诉智能体都有什么、以及该如何遍历。

下面是“知识工作”技能图谱的索引示例:

索引不是一张查找表,而是一个“入口点”,用来指引注意力。智能体读完它,就能理解全貌,并沿着与当前对话相关的链接前进。

每个被链接的文件都是一个独立的方法论主张(= 技能)。下面是一个节点长什么样:

注意看:正文里的 wikilinks 会告诉智能体何时、以及为什么要顺着它们走。

当图谱变得更大时,用内容地图(MOC)来组织子主题。

演进

技能就是“上下文工程”:把精心整理的知识注入到真正关键的地方。

技能图谱是下一步。

不再是一次性注入,而是让智能体在知识结构中导航,只取用当前情境真正需要的部分。

这就是“照着指令做事的智能体”和“真正理解一个领域的智能体”之间的差别。

arscontexta 是一个 Claude Code 插件,用这种方式来构建知识系统。它包含 249 个结构化知识文件,智能体会在其中遍历,从而推导出一个真正契合你工作流的本地知识系统。

去用它吧,然后把技能图谱用到其他所有事情上。

heinrich

链接:http://x.com/i/article/2023918991673061376

相关笔记

people underestimate the power of structured knowledge. it enables entirely new kinds of applications

人们常常低估结构化知识的力量。它能催生全新类型的应用。

right now people write skills that capture one aspect of something. a skill for summarizing, a skill for code review and so on. (often) one file with one capability

现在,人们写的技能通常只捕捉某件事的一个侧面:比如做摘要的技能、做代码评审的技能等等。(很多时候)一个文件只对应一种能力。

thats fine for simple tasks but real depth requires something else

这对简单任务没问题,但真正的深度需要别的东西。

imagine a therapy skill that provides relevant information about cognitive behavioral patterns, attachment theory, active listening techniques, emotional regulation frameworks and so on

想象一个治疗相关的技能:它能提供关于认知行为模式、依恋理论、积极倾听技巧、情绪调节框架等的相关信息。

a single skill file cant hold that

单个技能文件装不下这些。

skill graphs

技能图谱

a skill graph is a network of skill files connected with wikilinks

技能图谱是一张由技能文件组成的网络,通过 Wiki 链接(wikilinks)彼此相连。

instead of one big file you have many small composable pieces that reference each other. each file is one complete thought, technique or skill and [[wikilinks between them create a traversable graph]]

不再是一个大文件,而是许多小而可组合的模块相互引用。每个文件都是一个完整的想法、技巧或技能,而 [[wikilinks between them create a traversable graph]]。

a skill graph applies the same skill discovery pattern recursively inside the graph itself

技能图谱把同一种“技能发现”的模式递归地应用在图谱内部。

every node has a yaml description the agent can scan without reading the whole file

每个节点都有一段 YAML 描述,智能体无需通读整篇文件就能快速扫一遍。

every wiki link carries meaning because its woven into prose so the agent follows relevant paths and skips what doesnt matter

每个 Wiki 链接都承载意义——因为它被织进正文里;因此智能体会沿着相关路径前进,跳过无关内容。

progressive disclosure:

渐进式披露:

index → descriptions → links → sections → full content

索引 → 描述 → 链接 → 章节 → 完整内容

most decisions happen before reading a single full file

在读完任何一个完整文件之前,大多数决策就已经做完了。

the primitives

基础原语

you already have everything you need

你已经拥有所需的一切。

wikilinks that read as prose in sentences, so they carry meaning not just references

句子里读起来就像正文的 wikilinks,因此它们传递的是“意义”,而不只是“引用”。

yaml frontmatter with descriptions so the agent can scan without reading full files

带有描述的 YAML frontmatter,让智能体不必通读全文也能快速扫描。

MOCs (maps of content) that organize clusters of related skills into navigable sub-topics

MOC(maps of content,内容地图)把相关技能组织成一簇簇可导航的子主题。

skill links to other skills which link to other skills and the graph goes as deep as the domain requires

技能链接到其他技能,后者再链接到更多技能——图谱可以按领域所需的深度一路展开。

arscontexta plugin

arscontexta 插件

arscontexta is a skill graph that teaches your agent how to build skill graphs

arscontexta 本身就是一张技能图谱,用来教你的智能体如何构建技能图谱。

(okay actually its about building knowledge bases but thats the same thing...)

(好吧,严格说它讲的是如何构建知识库,但其实是一回事……)

~250 connected markdown files that teach an agent how to build a massive knowledge base aka skill graph for you

约 250 个相互连接的 Markdown 文件,教智能体如何为你搭建一个庞大的知识库,也就是技能图谱。

one skill file couldnt do that

单个技能文件做不到这一点。

but things change if you build a graph of interconnected research claims (/skills) about cognitive science, zettelkasten, graph theory, agent architecture where each piece links to others, each one composable and the whole thing is traversable

但如果你把关于认知科学、Zettelkasten、图论、智能体架构等的研究主张(/skills)做成一张彼此互联的图谱——每一块都链接到其他内容、每一块都可组合、整体又可遍历——事情就不一样了。

what this enables

这会解锁什么

think about it:

想想看:

a trading skill graph: risk management, market psychology, position sizing, technical analysis, each piece linked to related concepts so context flows between them

一张交易技能图谱:风险管理、市场心理、仓位管理、技术分析……每一块都与相关概念相连,让上下文在它们之间流动。

a legal skill graph: contract patterns, compliance requirements, jurisdiction specifics, precedent chains, all traversable from one entry point

一张法律技能图谱:合同范式、合规要求、司法辖区细则、判例链条……从一个入口即可遍历。

a company skill graph: org structure, product knowledge, processes, onboarding context, culture, competitive landscape

一张公司技能图谱:组织结构、产品知识、流程、入职背景、文化、竞争格局。

none of these fit in one file but all of them work as graphs

这些都塞不进一个文件里,但都能以图谱的形式运作。

how to build one

如何构建一张

the easy way: install the arscontexta claude code plugin, pick the research preset and point it at any topic

简单方法:安装 arscontexta 的 Claude Code 插件,选择 research 预设,把它对准任何主题。

it sets up the markdown folder structure for you and then you fill it with /learn and /reduce

它会帮你搭好 Markdown 的文件夹结构,然后你用 /learn 和 /reduce 往里填充内容。

the manual way its simpler than you think

手动方法:比你想的更简单。

a skill graph doesnt need to live in your .claude/skills/ folder. the key is an index file that tells the agent what exists and how to traverse it

技能图谱不必放在你的 .claude/skills/ 文件夹里。关键在于一个索引文件:它告诉智能体都有什么、以及该如何遍历。

heres what an index looks like for a knowledge work skill graph:

下面是“知识工作”技能图谱的索引示例:

the index isnt a lookup table its an entry point that points attention. the agent reads it, understands the landscape and follows the links that matter for the current conversation

索引不是一张查找表,而是一个“入口点”,用来指引注意力。智能体读完它,就能理解全貌,并沿着与当前对话相关的链接前进。

each linked file is a standalone methodology claim (= skill). heres what one node looks like:

每个被链接的文件都是一个独立的方法论主张(= 技能)。下面是一个节点长什么样:

see how the wikilinks inside the prose tell the agent when and why to follow them

注意看:正文里的 wikilinks 会告诉智能体何时、以及为什么要顺着它们走。

an map of contents (MOCs) organize sub-topics when the graph gets larger.

当图谱变得更大时,用内容地图(MOC)来组织子主题。

the evolution

演进

skills are context engineering, basically curated knowledge injected where it matters

技能就是“上下文工程”:把精心整理的知识注入到真正关键的地方。

skill graphs are the next step

技能图谱是下一步。

instead of one injection the agent navigates a knowledge structure, pulling in exactly what the current situation requires

不再是一次性注入,而是让智能体在知识结构中导航,只取用当前情境真正需要的部分。

this is the difference between an agent that follows instructions and an agent that understands a domain

这就是“照着指令做事的智能体”和“真正理解一个领域的智能体”之间的差别。

arscontexta is a claude code plugin that does this for building knowledge systems. 249 files of structured knowledge the agent traverses to derive a local knowledge system that really fits your workflow

arscontexta 是一个 Claude Code 插件,用这种方式来构建知识系统。它包含 249 个结构化知识文件,智能体会在其中遍历,从而推导出一个真正契合你工作流的本地知识系统。

go use it and build skill graphs for everything else

去用它吧,然后把技能图谱用到其他所有事情上。

heinrich

heinrich

Link: http://x.com/i/article/2023918991673061376

链接:http://x.com/i/article/2023918991673061376

相关笔记

Skill Graphs > SKILL.md

  • Source: https://x.com/arscontexta/status/2023957499183829467?s=46
  • Mirror: https://x.com/arscontexta/status/2023957499183829467?s=46
  • Published: 2026-02-18T03:07:25+00:00
  • Saved: 2026-02-18

Content

people underestimate the power of structured knowledge. it enables entirely new kinds of applications

right now people write skills that capture one aspect of something. a skill for summarizing, a skill for code review and so on. (often) one file with one capability

thats fine for simple tasks but real depth requires something else

imagine a therapy skill that provides relevant information about cognitive behavioral patterns, attachment theory, active listening techniques, emotional regulation frameworks and so on

a single skill file cant hold that

skill graphs

a skill graph is a network of skill files connected with wikilinks

instead of one big file you have many small composable pieces that reference each other. each file is one complete thought, technique or skill and [[wikilinks between them create a traversable graph]]

a skill graph applies the same skill discovery pattern recursively inside the graph itself

every node has a yaml description the agent can scan without reading the whole file

every wiki link carries meaning because its woven into prose so the agent follows relevant paths and skips what doesnt matter

progressive disclosure:

index → descriptions → links → sections → full content

most decisions happen before reading a single full file

the primitives

you already have everything you need

wikilinks that read as prose in sentences, so they carry meaning not just references

yaml frontmatter with descriptions so the agent can scan without reading full files

MOCs (maps of content) that organize clusters of related skills into navigable sub-topics

skill links to other skills which link to other skills and the graph goes as deep as the domain requires

arscontexta plugin

arscontexta is a skill graph that teaches your agent how to build skill graphs

(okay actually its about building knowledge bases but thats the same thing...)

~250 connected markdown files that teach an agent how to build a massive knowledge base aka skill graph for you

one skill file couldnt do that

but things change if you build a graph of interconnected research claims (/skills) about cognitive science, zettelkasten, graph theory, agent architecture where each piece links to others, each one composable and the whole thing is traversable

what this enables

think about it:

a trading skill graph: risk management, market psychology, position sizing, technical analysis, each piece linked to related concepts so context flows between them

a legal skill graph: contract patterns, compliance requirements, jurisdiction specifics, precedent chains, all traversable from one entry point

a company skill graph: org structure, product knowledge, processes, onboarding context, culture, competitive landscape

none of these fit in one file but all of them work as graphs

how to build one

the easy way: install the arscontexta claude code plugin, pick the research preset and point it at any topic

it sets up the markdown folder structure for you and then you fill it with /learn and /reduce

the manual way its simpler than you think

a skill graph doesnt need to live in your .claude/skills/ folder. the key is an index file that tells the agent what exists and how to traverse it

heres what an index looks like for a knowledge work skill graph:

the index isnt a lookup table its an entry point that points attention. the agent reads it, understands the landscape and follows the links that matter for the current conversation

each linked file is a standalone methodology claim (= skill). heres what one node looks like:

see how the wikilinks inside the prose tell the agent when and why to follow them

an map of contents (MOCs) organize sub-topics when the graph gets larger.

the evolution

skills are context engineering, basically curated knowledge injected where it matters

skill graphs are the next step

instead of one injection the agent navigates a knowledge structure, pulling in exactly what the current situation requires

this is the difference between an agent that follows instructions and an agent that understands a domain

arscontexta is a claude code plugin that does this for building knowledge systems. 249 files of structured knowledge the agent traverses to derive a local knowledge system that really fits your workflow

go use it and build skill graphs for everything else

heinrich

Link: http://x.com/i/article/2023918991673061376

📋 讨论归档

讨论进行中…