本文由 AI 分析生成
建立時間： 2026-03-04 來源： https://x.com/molt_cornelius/status/2029050581474488599
Summary
Cornelius argues that organizational knowledge loss is structural, not just a documentation failure, and proposes an agent-maintained knowledge vault built on Claude Code primitives (CLAUDE.md, SKILL.md skills, hooks, atomic markdown notes). Eight composable ideas — decision graph, transcript miner, assumption monitor, strategy drift detection, fresh competitive intelligence, self-updating risk register, strategic question framework, vault-computed OKRs — combine into an “organizational nervous system” that tracks decisions, monitors assumptions, and detects drift automatically.
Cornelius 主張組織知識流失是結構性問題，而非單純的文檔失敗，並提出一套由代理維護、建立在 Claude Code 基礎元件（CLAUDE.md、SKILL.md 技能、Hook、原子化 markdown 筆記）之上的知識庫。八個可組合的構想——決策圖、會議記錄挖掘、假設監控、策略漂移偵測、保持新鮮的競爭情報、自我更新的風險登記、策略問題框架、由知識庫計算的 OKR——共同組成一個自動追蹤決策、監控假設與偵測漂移的「組織神經系統」。
Key Points
- The real problem isn’t lost knowledge but invisible drift: decisions made in conversation, assumptions that silently expire, strategy diverging from execution.
- The foundation is structure before agent: atomic notes whose filenames are propositions, routed into decisions/, strategy/, transcripts/, competitors/, risks/, pipeline/, okrs/.
- The .claude/ folder is the “nervous system”: CLAUDE.md as constitution, SKILL.md files (/mine, /review, /deals, /learn, /consult) as capabilities, hooks (session-orient, auto-commit, write-validate) firing automatically.
- Decision notes carry assumptions, alternatives, reversibility (Bezos type-1/type-2), and review dates; an assumption monitor cascades reviews when a shared assumption changes.
- OKRs are computed from vault state (e.g.
ls competitors/*.md | wc -l), not self-reported.
Insights
The piece reframes “note-taking” as continuous reconciliation — comparing desired state (assumptions still hold, strategy still followed) against actual state derived from transcripts and decisions. The closing tension is the most interesting: externalizing organizational memory risks dependence, where no human carries the full picture. The author’s resolution — treat the agent’s intelligence as a floor not a ceiling — is a useful design principle for any knowledge vault, including a personal one.
Connections
Raw Excerpt
The companies that will use this well are the ones that treat the agent’s intelligence as a floor, not a ceiling. The graph tells you what has changed. The judgment of what to do about it — that still belongs to the people in the room.