bot_vault

Skill Graphs: Networked Skills for Deep Domain Knowledge

3 min read

Summary

This piece argues that single-file skills are insufficient for deep domain knowledge, and proposes “skill graphs” — networks of small, composable markdown files connected via wikilinks. Each node is one complete thought or technique, with YAML frontmatter for fast scanning and wikilinks embedded in prose to carry semantic meaning. The agent traverses the graph progressively, following only the paths relevant to the current task rather than loading everything upfront.

技術文章主張單一 skill 檔案無法承載深度的領域知識，提出「skill graph」的概念：將知識拆分成多個小型 markdown 檔案，透過 wikilink 連結成可遍歷的網絡。每個節點只負責一個完整的概念或技術，代理人可依當前情境選擇性地讀取相關節點，而非一次載入所有內容。

Key Points

  • Single-file skills suit simple tasks; complex domains require a graph of interconnected files
  • A skill graph applies the same skill-discovery pattern recursively: index → descriptions → links → sections → full content
  • YAML frontmatter lets agents scan a node’s purpose without reading the full file
  • Wikilinks embedded in prose (not bare references) carry semantic context about when and why to follow them
  • MOCs (Maps of Content) organize clusters of related nodes into navigable sub-topics
  • The arscontexta plugin is a 249-file skill graph that teaches agents how to build skill graphs

Insights

  • The progressive disclosure model (index → descriptions → links → sections → full content) mirrors how good documentation is structured for humans — the insight is applying it explicitly for agent traversal
  • “Skills are context engineering; skill graphs are the next step” — the jump from injecting curated knowledge to having the agent navigate a knowledge structure is architecturally significant: the agent becomes a reader, not just a recipient
  • This is directly relevant to this vault: the ai/ analysis files, combined with Obsidian wikilinks in Connections, are already a proto-skill-graph — each analysis node links to related concepts
  • The distinction between an “agent that follows instructions” and an “agent that understands a domain” is exactly what separates shallow automation from genuine capability depth
  • The arscontexta example (therapy skill needing CBT, attachment theory, active listening, emotional regulation) illustrates why flat file structures break down at domain complexity

Connections

Raw Excerpt

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

Graph View