karpathy-llm-wiki-boilerplate

A lightweight implementation of Andrej Karpathy's LLM Wiki approach — no engineering dependencies, just the minimal files needed to reproduce the core workflow. Clone it, configure it, and you'll have a personal Wiki continuously maintained by an LLM.

Core Philosophy

Traditional RAG derives everything from scratch on every query, with no knowledge accumulation. This approach is different:

Every time a source is ingested, the LLM compiles knowledge into the Wiki — updating summary pages, concept pages, entity pages, and cross-references. Knowledge compounds continuously rather than starting from zero each time.

Division of Labor: Humans curate sources, ask questions, and make decisions; LLMs handle summarization, archiving, cross-referencing, and consistency maintenance.

How the Workflow Works

This project's workflow is entirely driven by LLM + Prompt constraints — no code, databases, or external services needed. Four Markdown files define the operational rules, the LLM reads them and executes accordingly, and Git handles persistence.

The entire knowledge base is a Markdown repository: Humans feed sources, the LLM compiles and maintains, and Obsidian provides visual reading — clear division of labor with no interference.