Andrej Karpathy 的 LLM Wiki 工作流

 

Andrej Karpathy，OpenAI 共同創辦人、前 Tesla AI 負責人，最近分享了一個簡單卻深刻的想法：與其每次問 AI 問題時都從頭搜尋文件，不如讓 AI 把你的原始筆記「編譯」成一個結構化的知識庫，一次整理，持續更新。

他的做法是把所有資料存成純 Markdown 檔，分成三個資料夾：raw/ 放原始資料、wiki/ 讓 AI 自動生成概念頁面、output/ 存查詢結果。再搭配 Claude Code 這類終端工具，讓 AI 代理人自動讀取、整合、交叉連結所有筆記。理想狀態下，知識庫會隨著你加入新資料而自動成長，不需要人工維護。

聽起來很美好——但問題也很明顯：你要信任 AI 寫的東西，而 AI 寫的東西不一定值得信任。

---

你的作法：慢但更穩

你用自己寫的 Logseq 插件，手動匯出 1–2 層的連結筆記，合併成一個 Markdown 檔，再交給 LLM 產生概念索引。這個差異看似是「效率輸了」，但實際上你保留了幾個關鍵優勢：

輸入是你策展過的。 Karpathy 的系統需要自己判斷哪些內容相關；你的插件在交給 AI 之前，已經做了一次人工的圖遍歷，噪訊少很多。

輸出是你審閱過的。 AI 寫完你會讀，讀完才入庫。自動化系統的最大隱患是「AI 讀 AI 寫的東西」，錯誤會悄悄放大；你的流程每個環節都有人在迴路裡。

好奇心是你的過濾器。 自動化系統會把整個 Vault 都編譯一遍，產出一堆沒人查的頁面。你只整理你真正在乎的概念，索引頁的品質因此更高、更有個人意義。

代價是速度，以及需要你主動去觸發。但對於個人知識庫來說，「你真正信任且會使用的系統」永遠勝過「理論上更自動化但你不敢相信的系統」。

The Honest Tradeoffs

	Your approach	Full Karpathy
Trust in output	High — you review everything	Lower — AI writes, AI reads
Coverage	Curiosity-driven	Exhaustive (for better or worse)
Maintenance burden	Manual, but intentional	Script-dependent, can break
Compounding value	Grows with your interest	Grows with automation
Risk to existing vault	Near zero	Low but real

---

Prompt：產生概念索引

You are a knowledge base curator. I will give you a exported markdown file from my personal notes (Logseq/Obsidian), containing a central note and its 1-2 layers of linked notes merged into one document.

Your task is to produce a structured concept index page for the CENTRAL concept only. Do not summarize every note — synthesize them.

Output the following sections in Markdown:

## [Concept Name]
One concise paragraph: what this concept is, in my own words as implied by the notes. Avoid generic definitions — reflect how *I* seem to understand and use this concept.

## Core Ideas
3–6 bullet points capturing the most important insights from the notes. Each bullet should be a distinct idea, not a restatement.

## Key Relationships
List connections to other concepts in the format:
- [[concept A]] → relationship → [[concept B]]
Focus on meaningful relationships (causal, compositional, contrasting, enabling), not just co-occurrence.

## Tools & Resources
| Name | What it does | Link |
|------|-------------|------|
Only include tools/resources explicitly mentioned in the notes.

## Open Questions
Things the notes raise but don't resolve. Frame as genuine questions, not topics.

## Links Out
Other concepts or notes referenced in this export that deserve their own index page next. Ranked by how central they seem to this concept.

## Source Notes
List the note titles that contributed most to each section above. This is for auditability — so I know where each insight came from.

---
Here are my exported notes:

[PASTE YOUR EXPORTED MARKDOWN HERE]