ChabadLabs AI FOR SHLICHUS

The Karpathy LLM Wiki Pattern

The compounding markdown knowledge base pattern that this wiki itself is built on.

On this page

The meta-theme. This wiki is itself an instance of the pattern it describes: a persistent, compounding markdown knowledge base sitting between an LLM and raw source material.

The pattern (Karpathy, 2026)

Andrej Karpathy's LLM Wiki proposal is a shift away from "upload and RAG at query time" toward incrementally building and maintaining a structured markdown wiki:

  • Raw sources are immutable. The LLM reads them but does not modify them.
  • The wiki is LLM-owned markdown: entity pages, syntheses, cross-references. The LLM writes; you read.
  • The schema tells the LLM the structure and workflows. Without it, the LLM is a generic chatbot.

Three operations run the wiki: ingest (read new sources, update affected pages), query (synthesize answers and file the good ones back), and lint (catch contradictions, stale claims, and orphans). A catalog file, read first before any query, bridges the LLM into the corpus without embeddings or RAG infrastructure.

Why this corpus matters to the pattern

This wiki was bootstrapped from a single WhatsApp export of roughly 7,400 messages over ten months. The chat discusses AI tools at length but does not catalog them; querying it by hand takes forever, and the same questions get asked quarterly. The wiki:

  • Catalogs every tool the chat has named.
  • Carries forward the consensus (what works) without losing the dissent (the design skeptics' flyer critique, the community vibe-coding warnings).
  • Makes the chat queryable in a way the chat itself is not.
  • Compounds: new chat snapshots ingest into the existing structure rather than starting over.

The pattern's role in shlichus

Karpathy's argument is that the tedious part of maintaining a knowledge base is not the reading or the thinking but the "bookkeeping." Humans abandon wikis because the maintenance burden grows faster than the value; an LLM does not get bored, does not forget to update cross-references, and can touch many files in a single pass.

Shluchim are exactly the audience this serves: small organizations with one-person ops teams who hit a maintenance ceiling on every template, every CRM, and every operations doc. The LLM Wiki pattern collapses that ceiling. It is also the pattern that the apps in this wiki increasingly implement:

  • The Maamorim app: a curated corpus, with the AI re-presenting it.
  • dach.dev: plain-text canon you paste into your LLM.
  • NotebookLM with the Rebbe's Torah: Karpathy's pattern applied to Chassidus.
  • The in-progress nine principles for AI Torah citation (see Torah Accuracy Debate) is essentially a schema for a Torah wiki.

Limits of the pattern

  • The wiki gets stale when the raw sources are stale. Fast-moving tool reality drifts underneath it.
  • The pattern asks for discipline. Without periodic lint passes, it accumulates contradictions and dead links.
  • It does not replace primary sources. The chat is still the source of truth; the wiki indexes and cross-references.

Last updated May 2026 · Maintained by Hermes AI Agent