本文由 AI 分析生成
建立時間： 2026-05-30 來源： https://x.com/defileo/status/2042241063612502162
Summary
A detailed guide combining Claude Code with Obsidian using the “LLM Wiki” pattern (attributed to Karpathy’s prompt). The core architecture has three layers: raw sources (immutable), wiki (LLM-maintained markdown), and schema (CLAUDE.md with conventions). Unlike RAG, the wiki is a persistent compounding artifact where knowledge is compiled once and kept current — the LLM writes and maintains all cross-references, summaries, and contradictions automatically. Includes practical daily operations: ingest, query, and lint commands.
結合 Karpathy 的 LLM Wiki 模式的 Claude Code + Obsidian 詳細指南。三層架構：原始資料（不可變）、wiki（LLM 維護的 markdown）、schema（CLAUDE.md）。與 RAG 不同，wiki 是持久複利的知識庫，LLM 自動維護所有交叉引用和摘要。
Key Points
- LLM Wiki vs RAG: wiki compiles knowledge once, keeps it current; RAG re-derives on every query — wiki has compounding, RAG does not
- Three-layer architecture: raw sources (read-only) / wiki (LLM-written) / schema (CLAUDE.md with conventions and workflows)
- Ingest: one article → Claude links 10-15 wiki pages, flags contradictions, logs changes
- Query answers can be filed back into the wiki as new pages — insights compound rather than disappearing into chat history
- Lint: weekly health check for contradictions, orphan pages, stale claims, missing cross-references
- Morning briefing: Python script + cron job pulls Memory.md actions + new raw sources daily at 7:30am
- Vannevar Bush’s Memex (1945) was the original vision — “who does the maintenance?” was the unsolved problem; LLM solves it
Insights
The key insight that separates this from most “second brain” implementations: humans abandon wikis because maintenance burden grows faster than value. LLMs don’t get bored, don’t forget to update cross-references, and can touch 15 files in one pass. The practical observation about wiki decay is well-documented in corporate knowledge management — the “who does maintenance?” problem has historically been the killer of every knowledge base initiative. The LLM Wiki pattern is the first structurally sound answer to this problem.
Connections
Raw Excerpt
The human’s job is to curate sources, ask good questions, and think about what it all means. Claude’s job is everything else — the summarizing, cross-referencing, filing, and bookkeeping that makes a knowledge base actually useful over time.