Descriptions:
Andrej Karpathy recently shared a post about using LLMs to build personal knowledge bases from raw documents, and this video from Nate Herk translates that idea into a working system anyone can replicate in about five minutes using Claude Code and Obsidian. The result is a self-organizing markdown vault where Claude Code ingests raw source material — YouTube transcripts, articles, PDFs — and automatically extracts entities, concepts, and relationships without any manual tagging.
The video shows two live examples: a YouTube knowledge system organizing 36 of Nate’s own videos into a graph of tools, techniques, and backlinks; and a personal second brain tracking business initiatives, employees, and quarterly goals. The core architecture is deliberately simple — a raw folder for source documents and a wiki folder where Claude Code writes structured markdown — following Karpathy’s own observation that fancy vector databases and embeddings aren’t necessary at this scale (he runs roughly 100 articles and 500,000 words on the same pattern).
Practical tooling covered includes the Obsidian Web Clipper Chrome extension for clipping articles directly into the vault, Claude Code’s CLAUDE.md schema for instructing the wiki agent, and the tradeoffs between flat and nested folder structures. The key insight demonstrated throughout is that because Claude Code maintains index files and brief summaries automatically, querying the knowledge base becomes fast and context-rich — making it a viable long-term memory layer for AI-assisted workflows.
📺 Source: Nate Herk | AI Automation · Published April 05, 2026
🏷️ Format: Tutorial Demo
