personal-wiki/wiki/sources/Nate Herk LLM Wiki Transcript.md

type, title, source_type, author, date_published, url, confidence, key_claims, created, updated, tags, status, related, sources

type	title	source_type	author	date_published	url	confidence	key_claims	created	updated	tags	status	related	sources
source	Nate Herk LLM Wiki Transcript	transcript	Nate Herk	2026-04-07	https://youtube.com/@nateherk	high	LLM wiki makes knowledge compound like interest — nothing is re-derived on every query Hot cache (~500 words) enables cross-project context without crawling the full wiki One article can generate 15-25 wiki pages with full cross-references One user dropped token usage by 95% switching from inline context files to wiki Obsidian is the IDE, Claude is the programmer, the wiki is the codebase Index file is enough at small scale (~100 sources) — no RAG infrastructure needed	2026-04-07	2026-04-07	source llm-wiki obsidian karpathy	mature	LLM Wiki Pattern Hot Cache Compounding Knowledge Andrej Karpathy index	.raw/nate-herk-llm-wiki-transcript.md

Nate Herk LLM Wiki Transcript

Raw source: .raw/nate-herk-llm-wiki-transcript.md

Nate Herk demonstrates the LLM Wiki Pattern in practice. He shows two live vaults: one for his YouTube transcript archive (36 videos) and one personal second brain. He breaks down Andrej Karpathy's original post and shows a 5-minute setup workflow.

---

Key Takeaways

The core insight: normal AI chats are ephemeral. The wiki makes knowledge compound. Every source ingested, every question answered, every analysis filed — all of it stays and grows richer over time.

The stack is simple: Claude Code + Obsidian + a folder of markdown files. No vector databases, no embeddings, no infrastructure. Just files and Claude.

The hot cache: a ~500-word file (wiki/hot.md) that captures recent context. In an executive assistant setup, this prevented having to crawl dozens of wiki pages at the start of each session. See Hot Cache.

Cross-project referencing: other Claude Code projects can read this vault by pointing at it in their CLAUDE.md. Nate's executive assistant reads from his herk-brain vault. Token usage dropped significantly compared to inline context files.

At scale: the index file alone is sufficient for hundreds of pages. Vector RAG only becomes necessary at millions of documents.

---

Obsidian as IDE

Obsidian is just a markdown viewer with graph visualization. The graph view shows which pages are hubs (many connections) and which are orphans (none). Real-time — you can watch the wiki grow as Claude creates pages.

The key Obsidian features used:

• Graph view — visualize the knowledge structure
• Backlinks — follow connections between pages
• Dataview — query pages by frontmatter
• Web Clipper — send articles directly to .raw/ from any browser

---

Workflow Demonstrated

1. Install Obsidian, create a vault
2. Paste Karpathy's LLM wiki idea into Claude Code
3. Claude scaffolds the structure (raw/, wiki/, CLAUDE.md, index, log)
4. Drop a source into .raw/ using Web Clipper
5. Tell Claude: "ingest this"
6. Claude reads, creates 15-25 wiki pages, cross-references everything
7. Query the wiki for insights

The ingest for one article (AI 2027) took 10 minutes and created 23 pages: 1 source, 6 people, 5 organizations, 1 AI systems page, multiple concepts, plus an analysis.

---

Entities Mentioned

• Andrej Karpathy — originated the LLM wiki pattern
• Nate Herk — demonstrated the pattern in this video

---

Connections

See LLM Wiki Pattern for the full architecture. See Compounding Knowledge for the core insight on why this works. See Hot Cache for the session context mechanism.