Smorster

Andrej Karpathy Just 10x’d Everyone’s Claude Code

Nate Herk | AI Automation · 17m 47s · Watch on YouTube · 10 sources

Decision Card

Effort: Weekend project — download Obsidian (free), create a vault with raw/ and wiki/ folders, paste Karpathy’s gist into Claude Code, and ingest your first PDF/article; expect ~10 minutes per source on the LLM’s side.

Honest take: The “5 minutes” framing is misleading — setup is fast, but each ingest takes ~10–14 minutes of Claude Code working, and the video’s own closing admission is the real catch: this pattern only works at small scale (hundreds of pages), and breaks down into needing real RAG once you hit millions of documents. It’s a genuinely useful pattern repackaged as a “10x” hype headline; Karpathy’s name is doing heavy lifting since the actual idea is “organized markdown files.”

Concrete next steps:

  • Read the primary source directly — Karpathy’s LLM wiki gist (5 min) — before watching any tutorial; the video adds little beyond it.
  • Install the official Obsidian Web Clipper and set its save location to your raw/ folder (10 min).
  • Stand up a throwaway vault, ingest 2–3 sources, then run a “lint” pass to see whether the auto-generated structure actually fits your material (1–2 hrs incl. ingest wait).
  • Skip if you have more than a few thousand documents, need multi-user/enterprise scale, or already have a working RAG/semantic-search pipeline — the video itself concedes this pattern doesn’t scale there.

TL;DR

The video walks through Andrej Karpathy’s April 2026 “LLM wiki” idea: instead of RAG, you let Claude Code ingest raw documents and build a structured, interlinked folder of markdown files that compounds over time, optionally visualized in Obsidian. It’s cheap and simple at the scale of ~100 sources, but the creator openly admits it doesn’t scale to enterprise/millions-of-documents use.

Key Points

  • The creator turned 36 of his YouTube transcripts into an auto-organized, backlinked knowledge graph just by telling Claude Code to grab the transcripts and “organize everything” — no manual relationship building 00:31
  • The idea comes from an Andrej Karpathy X post about using LLMs to build personal knowledge bases, which went viral within days 01:41
  • The workflow has three stages: data ingest (drop source docs into Claude Code), organization into a wiki, and a Q&A phase where you query the whole wiki 02:03
  • Karpathy said he thought he’d need “fancy RAG,” but the LLM auto-maintains index files and brief summaries and reads related data fine at small scale (~100 articles, ~half a million words) 02:43
  • No vector database, embeddings, or complex infrastructure needed — “it’s literally just a folder with markdown files” with a raw folder, a wiki folder, an index, and a log 03:37
  • One X user reportedly turned 383 scattered files and 100+ meeting transcripts into a compact wiki and dropped query token usage by 95% 04:53
  • Obsidian (free) is used only as a visual front-end for the markdown; the Obsidian Web Clipper extension pulls web articles straight into the raw folder 06:08
  • Karpathy suggests periodic “linting” — LLM health checks to find inconsistent data, impute missing data via web search, and surface new article candidates 15:04
  • An optional hot.md cache stores ~500 recent words to save crawling the full wiki — useful for an executive-assistant use case, unnecessary for the YouTube one 14:42
  • The honest caveat: the wiki approach wins at hundreds of pages with good indexes, but at millions of documents you’ll still want traditional RAG, a knowledge graph, or LightRAG 17:09

Notable Quotes

“I literally didn’t have to do any manual relationship building here. It just figured it all out on its own.” 01:07

“You don’t need a fancy vector database, embeddings, or complex infrastructure. It’s literally just a folder with markdown files.” 03:38

“If you have hundreds of pages with good indexes, you’re fine with wiki graph. But if you were getting up to the millions of documents, then you’re going to want to actually do more of a traditional rag pipeline.” 17:10

Verified Claims

Karpathy posted about LLMs building personal knowledge bases, and it went viral in April 2026. 01:41

Karpathy said he thought he’d reach for fancy RAG, but the LLM auto-maintains index files and summaries well at small scale. 02:43

  • Karpathy’s llm-wiki gist — notes the index approach “works surprisingly well at moderate scale (~100 sources, ~hundreds of pages) and avoids the need for embedding-based RAG infrastructure”
  • Verdict: Confirmed

The scale described is ~100 articles and ~half a million words. 02:54

The workflow needs no vector database/embeddings — just a folder of markdown files with raw/wiki/index/log. 03:38

  • Karpathy’s llm-wiki gist — describes a raw/ sources layer, an LLM-owned wiki/ layer, and an index.md catalog
  • Verdict: Confirmed

Karpathy recommends periodic “linting” health checks (find contradictions, impute missing data via web search, surface new connections). 15:04

  • Karpathy’s llm-wiki gist — the “Lint” stage health-checks for contradictions, stale claims, orphan pages, and data gaps fillable by web search
  • Verdict: Confirmed

Obsidian is free, and the Web Clipper is an official extension that saves web pages as markdown. 06:08

An X user dropped query token usage by ~95% by converting 383 files / 100+ transcripts into a wiki. 04:53

The wiki pattern doesn’t scale to millions of documents — traditional RAG/knowledge graphs are needed there. 17:09

Tools, Papers & Standards Mentioned

  • Andrej Karpathy’s llm-wiki gist — the canonical idea file
  • Claude Code — the agent that ingests and maintains the wiki
  • Obsidian — free markdown editor / graph-view front-end
  • Obsidian Web Clipper — official browser extension for clipping web pages to markdown (GitHub)
  • AI 2027 — the article ingested in the demo
  • LightRAG — referenced as a scaling alternative for large corpora
  • CLAUDE.md — schema/instructions file (referred to in the video as “claw.md”)

Follow-up Questions

  1. At what real-world document count and token budget does the markdown-wiki approach actually become more expensive than embedding-based RAG, and can that crossover be measured rather than asserted?
  2. How reliable is the LLM at maintaining cross-reference integrity over many ingest cycles — does the “lint” pass reliably catch contradictions and orphan pages, or does drift accumulate?
  3. Can the same wiki vault be safely shared across multiple agents/projects (the “point another project at this folder” claim) without one agent corrupting another’s structure or the index?

Sources