本文由 AI 分析生成
建立時間： 2026-05-30 來源： https://x.com/arscontexta/status/2023957499183829467
Summary
The post introduces “skill graphs” — networks of interconnected markdown skill files connected via wikilinks — as a more powerful alternative to single-file skills for agents. Each node is a complete thought/technique with YAML frontmatter for scanning, and wikilinks carry semantic meaning so agents can traverse relevant paths. The progressive disclosure pattern (index → descriptions → links → sections → full content) means most decisions happen before reading any full file.
這篇文章提出「skill graph」概念：將 agent 的知識組織成相互連結的 markdown 檔案網絡，透過 wikilinks 建立可遍歷的圖結構。與單一技能文件不同，圖結構允許深度領域知識。核心是漸進式揭露模式：agent 從索引掃描到描述再到完整內容，大多數決策在讀完整文件之前就已做出。
Key Points
- Single skill file can’t hold complex domain knowledge (therapy, legal, trading)
- Skill graph: many small composable files with meaningful wikilinks
- YAML frontmatter allows scanning without reading full content
- Progressive disclosure: index → descriptions → links → sections → full content
- MOCs (Maps of Content) organize clusters into navigable sub-topics
- arscontexta: 250 connected files teaching agents to build knowledge graphs
Insights
The key insight is that wikilinks embedded in prose carry meaning (the agent knows why to follow a link), unlike bare reference lists. This mirrors good zettelkasten practice but adapted for agent traversal rather than human reading. The recursive quality — a skill graph that teaches you to build skill graphs — is elegant. The distinction from rules files (always loaded) vs. skills (selectively loaded) maps directly to context window economics.
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.”