How to Build a Personal LLM Knowledge Base (Karpathy’s Method)

Main concept: “Personal LLM knowledge base” (AI second brain)

The video describes a system popularized by AI figures like Andrej Karpathy and Lex Fridman. The idea is to build an AI-maintained wiki from sources you choose.

The resulting wiki:

• Grows over time
• Lets you ask complex questions
• Can surface agreements and disagreements across sources
• Acts like an “LLM that doesn’t forget,” because it’s grounded in your collected, structured notes

---

High-level workflow (3 phases)

1) Collect

• Gather the best content on a topic.
  • Example topic used: hypertrophy / building muscle
• Sources mentioned:
  • YouTube transcripts
  • Substack
  • Podcast transcript research summaries
• Save everything locally into a folder using Obsidian Web Clipper.

YouTube tip: use the built-in transcript from the video description, copy it, and save it as a text file in the “raw” folder.

---

2) Compile

• Use Claude (via Claude Code / “Cloud Cowork” mentioned) to read everything in the raw folder.
• Provide a prompt that instructs Claude to generate a structured wiki, including:
  • Interconnected Markdown files
  • Organized by concept, not by source/person
  • Each concept becomes a standalone article that:
    • summarizes evidence
    • flags where sources disagree (and why)
    • links related articles
  • A generated index.md containing every article for fast navigation

Example concept pages from the demo:

• body recomposition
• calorie & surplus
• hypertrophy mechanisms
• strength vs hypertrophy
• sleep & recovery

---

3) Query

• Point Claude at the compiled wiki and ask higher-level questions, such as:
  • “What’s the most important variable for hypertrophy?”
  • “How do I optimize it?”
  • “Where do my sources agree vs disagree?”
  • “What gaps exist in my wiki / what should I learn next?”

The video emphasizes that once the wiki + index exist, querying becomes more token-efficient, because you avoid re-reading raw sources repeatedly.

Example demo conclusions for hypertrophy:

• Mechanical tension is the primary driver (agreed upon by the sources)
• Progressive overload is the “meta principle” that helps ensure tension increases over time

Example “gap” findings from the demo:

• strong on fundamentals
• weaker on application troubleshooting and specialization
• missing items like a decision tree for debugging stalled progress, plus more on deloading and advanced intensity techniques (drop sets, rest-pause, myo-reps, supersets, etc.)

---

Tutorial / product details (how to build it)

Tools and setup

• Download and set up Obsidian
  • Create a new Vault
  • Use a common naming convention: a “raw” folder for input sources
• Install the Obsidian Web Clipper
  • Save webpages into your local vault

Compilation

• Compile using Claude Code
  • Requires a paid Claude tool: $20/month mentioned
  • Demo timing: ~20 minutes (and ~32.1k tokens used)

Use the output

• Open the generated wiki in Obsidian
  • Review the Markdown concept files
  • Use Graph view to visualize connections between concepts (described as visually cool, though not always most useful depending on preference)

Key idea: once compiled, the wiki becomes your navigable, queryable structured “knowledge base.”

---

Notable real-world usage mentioned

• Lex Friedman reportedly:
  • Fed this kind of personal wiki into voice mode
  • Uses it while running to ask questions and explain things
  • Uses it similarly to the Feynman technique—identifying what he doesn’t understand by comparing his explanation to what the wiki says

---

Main speakers / sources mentioned

• Andrej Karpathy (initiated/popularized the idea; referenced via tweet)
• Lex Fridman (claimed to do the same; described the voice-mode use case)
• Dr. Andy Galpin (example expert source used via a Huberman Lab podcast transcript)
• Huberman Lab (podcast source referenced indirectly)