本文由 AI 分析生成
建立時間： 2026-05-25 來源： https://x.com/FarzaTV/status/2040563939797504467
Summary
Farza fed 2,500 personal diary entries, Apple Notes, and iMessages to an LLM, which generated 400 backlinked Wikipedia-style articles organized around people, startups, research areas, and interests. The resulting “Farzapedia” was not built for human browsing — it was deliberately structured for agent crawlability (starting from index.md, drilling into specific articles on demand). Inspired by Karpathy’s idea file concept, the system allows an AI agent to retrieve personal context on demand and continuously self-updates as new material is added.
將 2,500 條個人日記、筆記和 iMessage 輸入 LLM，生成 400 篇有回鏈的維基百科式文章。「Farzapedia」不是為人類瀏覽設計的，而是為代理人可爬取性設計的——從 index.md 出發、按需鑽取特定頁面，並在新內容加入時自動更新相關文章。
Key Points
- 2,500 personal entries → 400 detailed articles with backlinks across friends, startups, research areas, creative influences
- Key design principle: structured for agent crawlability, not human browsing; file system structure + backlinks beat RAG for agent retrieval
- Agent use: spin up Claude Code on the wiki, start at index.md, agent drills into relevant pages as needed for each query
- Example query: “look into images and films that inspired me recently and give ideas for copy and aesthetics” → agent pulls Philosophy articles from Ghibli notes, Competitor articles with YC screenshots, 1970s Beatles merch images
- System self-updates: new additions automatically update 2-3 related articles or create new ones
- RAG comparison: author tried RAG a year earlier and found it inferior — structured file systems that agents understand outperform vector similarity for this use case
Insights
The “built for the agent, not for me” reframe is the key insight: most PKM systems optimize for human recall (tags, folders, search), but agent-accessible knowledge systems optimize for machine traversal (consistent structure, explicit backlinks, master index). The flat file system with wikilinks is effectively a graph database that both humans and agents can navigate. This is a concrete implementation of the “idea file” Karpathy described — sharing the idea (structured personal wiki) rather than the specific code, so any agent can rebuild the capability.
Connections
Raw Excerpt
A knowledge base that lets an agent find what it needs via a file system it actually understands just works better.