Heidelberg AICurriculum

Fable 5 + Karpathy’s LLM Wiki is Basically Cheating

Nate Herk | AI Automation ·2026-07-03 ·1 min read

Summary written by us from the video's transcript. The video, and everything in it, is Nate Herk | AI Automation's work.

Learn how to quickly set up an LLM‑powered wiki that ingests YouTube transcripts, PDFs, or URLs and creates a navigable second brain of linked markdown files.

Takeaways

  • A vault with raw and wiki folders lets Claude automatically turn videos, PDFs, or URLs into linked markdown pages.
  • One prompt can ingest multiple sources; Claude creates index entries, logs, and cross‑links without manual wiring.
  • Backlinks enable navigation between concepts (e.g., GitHub ↔ Vercel) just like a personal knowledge graph.
  • The folder structure (flat vs. hierarchical) is configurable to suit different data types such as meeting transcripts or YouTube videos.
  • Because the wiki consists of plain markdown files, any LLM‑enabled tool can read or write it, making the second brain portable.

Creating the Obsidian vault

Open obsidian.md, run the setup wizard for your OS, and click Manage Vaults → New Vault. Name it (e.g., AI test) and choose a location such as your desktop or a project folder like Herc 2. This creates an empty vault with a default welcome.md file.

Connecting Claude code to the vault

Open the vault folder in VS Code, locate the .obsidian directory, then launch Claude (or any Claude‑compatible interface) from the terminal. The model will now have read/write access to the vault’s markdown files.

Preparing the raw data folder

Inside the vault create a raw folder for source files and a wiki folder where processed pages will live. Raw items can be PDFs, URLs, or plain text that Claude will ingest.

Ingesting sources with a single prompt

Drop a PDF (e.g., Claude Fable 5 system card) into the raw folder. For a web article, paste its URL in the prompt: “Read this article and ingest it into our wiki.” Include the PDF path as well. Claude will parse each source, create an index entry, log the operation, and generate linked markdown pages under concepts, entities, sources, etc.

Exploring the generated wiki

After ingestion you’ll see dozens of cross‑linked pages: tools, techniques, models, and their relationships (e.g., OpenAI GPT‑5.6 Soul ↔ Mythos 5). The index lists top‑level categories; the log records each batch ingest; backlinks let you jump from a model name to related concepts or source videos.

Iterating and customizing structure

If the automatic folder layout isn’t ideal, edit the schema (e.g., flatten vs. hierarchical) and re‑run ingests. Adjust routing rules in Claude so future documents are placed in the desired subfolders. The system works with any markdown‑based agent because everything is plain .md files.

Next steps and resources

Follow the free “Build your own AIOS” course linked in the description for a step‑by‑step walkthrough. You can also replicate Karpathy’s LLM wiki gist by copying it into Claude and prompting it to create the full schema.