Why Sharing Ideas Beats Sharing Code: Karpathy’s LLM‑Powered Wiki Workflow

One‑sentence overview

Karpathy’s core idea is to let an LLM pre‑compile raw materials into a structured, self‑maintaining Wiki so that users query the Wiki instead of re‑reading original documents.

Three‑layer architecture

The system consists of three layers:

Raw sources : immutable original documents such as papers, blogs, code repositories and datasets.

The Wiki : markdown files generated by the LLM, interlinked and cross‑referenced.

The Schema : rule files (e.g., CLAUDE.md, AGENTS.md) that tell the LLM how to organize and maintain the Wiki, including directory structure and naming conventions.

Three core operations

With the three layers in place, the workflow repeatedly performs three operations:

Ingest : When new material is added, the LLM reads it and updates 10‑15 wiki pages—concept pages, summaries, cross‑links, and contradictions.

Query : Users ask complex questions; the LLM answers from the structured knowledge instead of retrieving raw fragments, and the answer can be fed back to improve the Wiki.

Lint : A periodic health‑check that finds contradictions, missing links, or isolated pages, and suggests further exploration.

Difference from RAG

Traditional Retrieval‑Augmented Generation (RAG) retrieves raw fragments at query time and stitches an answer, relying on chunking and vector similarity. Karpathy’s approach compiles the raw material into a structured Wiki beforehand, so queries read from an already organized knowledge base, eliminating the need for on‑the‑fly retrieval.

He notes that for a knowledge base of about 100 articles (≈400 k words), the LLM‑maintained index and summaries are sufficient, and the system can accurately locate relevant content without a vector database.

Toolchain

The lightweight toolchain includes:

Obsidian : IDE for viewing raw data, the generated Wiki, and charts.

Obsidian Web Clipper : Converts web articles to markdown.

Marp : Renders markdown slides.

Custom search engine : Built with vibe coding, callable from CLI and usable as an LLM tool.

Karpathy emphasizes that the Wiki content is almost entirely written by the LLM; the human only feeds material and asks questions.

Actual results

In his own research domain, the Wiki has grown to roughly 100 articles and 400 k words. He can pose complex questions, and the LLM retrieves, synthesizes, and returns answers, sometimes generating markdown, slides, or matplotlib charts, which are then archived back into the Wiki, further enriching it.

How to replicate

The process can be reproduced in four steps:

Open Karpathy’s Gist: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

Copy the entire content (or give the link directly to your agent).

Tell your agent: “Build an LLM Wiki system for my use case based on this specification.”

The agent will ask about your scenario (research, reading, work) and generate the directory structure, schema files, and basic tool flow.

If you use the OpenClaw agent, simply provide the Gist link and it will construct the system without you writing code.

This approach turns knowledge management into a self‑growing, self‑repairing system, reducing reliance on personal discipline and manual curation.