Andrej Karpathy Just 10x'd Everyone's Claude Code¶
Summary¶
Nate Herk demonstrates building LLM-powered knowledge bases using Obsidian and Claude Code, showing two real wikis: one for 36 YouTube video transcripts and one for his personal second brain. Compares LLM Wiki to traditional semantic search RAG.
Key Takeaways¶
Real-World Wiki Examples¶
YouTube Transcript Wiki — 36 recent videos organized into a knowledge system: - Nodes for tools (Claude Code, Perplexity, VS Code), techniques (WAT framework, bypass permissions mode, human review checkpoint), MCP servers, RAG, vibe coding - AI automatically grabbed transcripts and organized everything — no manual relationship building - Backlinks connect videos to concepts, tools, and techniques mentioned
Personal Second Brain ("Herk Brain") — personal life, business (UpptAI), YouTube channel, employees, Q2 initiatives
Karpathy's Original Gist¶
Karpathy's gist (github.com/karpathy/442a6bf555914893e9891c11519de94f) describes the stages: 1. Data ingest — drop source documents (PDFs, articles) into the system 2. Obsidian as IDE — visually see markdown files and relationships 3. Q&A phase — ask questions, LLM reads through the entire wiki efficiently 4. Gap identification — LLM identifies gaps, does research, fills them
Karpathy's wiki at the time: ~100 articles, ~500,000 words. Left the prompt vague intentionally so people can customize it.
Token Efficiency¶
One X user turned 383 scattered files and 100+ meeting transcripts into a compact wiki, dropping token usage by 95% when querying with Claude.
Ingest Walkthrough¶
- Used Obsidian Web Clipper to save an article (AI 2027) directly into
raw/ - Told Claude Code to ingest; it read the article, determined it needed ~25 wiki pages
- Ingest took ~10 minutes; created 23 wiki pages: source, 6 people, 5 organizations, 1 AI systems page, concepts (technical alignment, geopolitical), analysis
- Graph view showed hubs (major connected nodes) vs. individual nodes in real time
Flat vs. Structured Wiki¶
Karpathy noted: "Sometimes I like to keep it really simple and really flat" — no subfolders, no over-organizing. Cole's YouTube wiki used subfolders (analysis, concepts, entities, sources) which made more sense for that use case.
LLM Wiki vs. Traditional RAG Comparison¶
| Dimension | LLM Wiki | Semantic Search RAG |
|---|---|---|
| Discovery | Reads indexes, follows links | Similarity search over vectors |
| Understanding | Deep relationships via links | Chunk similarity |
| Infrastructure | Just markdown files | Embedding model, vector DB, chunking pipeline |
| Cost | Basically free (tokens only) | Ongoing compute and storage |
| Maintenance | Lint, clean up, add articles | Re-embed when things change |
| Scale limit | Hundreds of pages (fine with good indexes) | Millions of documents |
Hot Cache (hot.md)¶
A ~500-word/character cache of the most recent conversation/context. Useful for executive assistants (Nate's "Herk 2") that need quick context without crawling full wiki pages. Not needed for all wiki types.
Pointing Other Agents to the Wiki¶
You can point a different project at the wiki folder — the agent crawls through the wiki, reads the index, and understands the project via AGENTS.md/CLAUDE.md. Nate configured his executive assistant with a wiki_path pointing to his Herk Brain vault.
When to Use Each¶
- Hundreds of pages with good indexes → Wiki graph is fine
- Millions of documents → Traditional RAG pipeline needed (as of April 2026)