Knowledge Wiki

yt-llm-wiki-knowledge-system

2 min read

YT — LLM Wiki Knowledge System

Type: source Tags: llm-wiki, hot-cache, token-efficiency, ingest-protocol, lint-protocol, andrej-karpathy, claude-code, obsidian, obsidian-web-clipper, llm-wiki-vs-rag Source type: YouTube transcript File: docs/transcript.md Approx date: April 2026

Summary

A YouTube walkthrough demonstrating how to build a personal knowledge system in Obsidian using the Karpathy LLM wiki concept. The presenter shows how Claude Code can automatically ingest raw source material (transcripts, articles) and organize it into an interconnected wiki of markdown files — no vector DB needed.

Key Takeaways

  • Drop source files into raw/, ask Claude Code to ingest → wiki pages created automatically with backlinks
  • Claude navigates via index files and [[wikilinks]], not embeddings
  • One user compressed 383 files + 100 meeting transcripts into a wiki and dropped token usage by 95%
  • Hot cache (hot.md) reduces token cost for repeat queries in executive-assistant workflows
  • Lint protocol lets Claude identify gaps, broken links, and suggest new articles
  • Scale limit: hundreds of pages (use RAG for millions)

Notable Quotes

“I thought that I had to reach for fancy RAG, but the LLM has been pretty good about automaintaining index files and brief summaries of all documents.” — Andrej Karpathy

“Normal AI chats are ephemeral. This method makes knowledge compound like interest in a bank.”

Workflow Shown

  1. Create Obsidian vault → open in Claude Code
  2. Paste Karpathy’s idea prompt → Claude bootstraps raw/, wiki/, CLAUDE.md
  3. Use Obsidian Web Clipper to drop articles into raw/
  4. Tell Claude: “ingest [filename]” → 25 wiki pages created from AI 2027 article (took ~10 min)
  5. Query the wiki in natural language → Claude reads index → follows links → synthesizes
  6. Optionally connect wiki to another Claude project (e.g., executive assistant) via wiki path in CLAUDE.md

Example Ingest Result (AI 2027 article)

  • 23 wiki pages created from one article
  • 1 source page, 6 people, 5 organizations, 1 AI systems page, several concepts, 1 analysis

Tools Mentioned

  • claude-code — runs ingest, creates wiki pages, answers queries
  • obsidian — visualizes markdown graph; optional but useful for seeing relationships
  • obsidian-web-clipper — Chrome extension for clipping web articles directly to raw/

Techniques Referenced (not detailed in this source)

  • WAT Framework
  • Bypass Permissions Mode
  • Human Review Checkpoint

Graph View

Backlinks