Andrej Karpathy Just 10x’d Everyone’s Claude Code
Source: YouTube video by Nate Herk (published 2026-04-05) Raw: Original Transcript
Summary
Nate Herk walks through Andrej Karpathy’s viral LLM Wiki idea and demonstrates how to implement it in ~5 minutes using Obsidian and Claude Code. The core pattern: drop raw source documents into a folder, let Claude Code read and organize them into a structured, interlinked wiki of markdown files.
Herk shows two working implementations:
- YouTube Transcript Wiki — 36 videos ingested in ~14 minutes, producing entity pages for tools, techniques, people, and concepts with full cross-references
- Personal Second Brain — meeting recordings, business notes, ClickUp summaries organized into a queryable wiki that his AI executive assistant can reference
Key Details
- Setup: Create an Obsidian vault with
raw/andwiki/folders. Paste Karpathy’s idea file into Claude Code. Done. - Ingesting: Use Obsidian Web Clipper to clip articles into
raw/. Tell Claude Code to ingest. A single article (AI 2027) generated 23 wiki pages in ~10 minutes. - Querying: Ask questions against the wiki. Claude reads the index, follows links, synthesizes answers. Valuable answers can be filed back as wiki pages.
- Linting: Periodic health checks — find contradictions, stale data, orphan pages, gaps. The LLM can suggest new sources to fill gaps.
- Hot Cache: A small file (~500 words) summarizing the most recent context. Useful when another agent needs quick access without crawling the full wiki.
- Cross-project use: The wiki directory can be referenced by other Claude Code projects (e.g., an executive assistant) by pointing to the wiki path in their
CLAUDE.md.
LLM Wiki vs Traditional RAG
| Aspect | LLM Wiki | Semantic Search RAG |
|---|---|---|
| Search method | Index files + wikilinks | Embedding similarity |
| Infrastructure | Markdown files only | Embedding model + vector DB + chunking pipeline |
| Cost | Token usage only | Ongoing compute + storage |
| Maintenance | Lint + add sources | Re-embed on changes |
| Scale | Hundreds of pages | Millions of documents |
| Depth | Deep relationships via links | Surface-level chunk similarity |
The wiki approach wins for personal/small-scale knowledge bases. Traditional RAG is needed at enterprise scale (millions of documents).
Notable Claims
- One X user reduced token usage by 95% after converting 383 scattered files + 100 meeting transcripts into a wiki.
- Karpathy intentionally left the idea vague so users could customize it to their domain.
- Karpathy manages ~100 articles / ~500K words with this approach, using index files instead of embedding-based retrieval.