Back to directory

karpathy-llm-wiki

An installable Agent Skills package that turns Claude Code, Cursor or Codex into a maintainer of a Karpathy-style LLM wiki: compiling ingested sources into durable, cross-linked markdown pages and answering questions with citations.

karpathy-llm-wiki

karpathy-llm-wiki packages the idea of an LLM-maintained wiki, where a model keeps structured knowledge pages instead of re-searching raw sources on every question, into one installable Agent Skills package for Claude Code, Cursor, Codex and other Agent Skills-compatible tools. It supports three operations: ingesting a new source, triaging it and creating or updating the relevant wiki pages, or simply logging it if nothing new was found; querying the wiki and answering with citations back to the pages that contain the synthesized knowledge; and linting the wiki for index integrity, broken links and overall health, with automatic fixes where possible.

It contrasts this approach with retrieval-augmented generation: an LLM wiki keeps knowledge in curated markdown pages rather than raw chunks and embeddings, and synthesizes that knowledge during ingestion and maintenance rather than at query time, which the project argues suits compounding, cross-linked knowledge better than broad retrieval across a large raw corpus.

A production knowledge base maintained with the skill since April 2026 is reported to hold 94 wiki articles across 13 topic directories from 99 ingested sources. It installs with a single command and is released under the MIT license.

More in Agent skills

See all 581
  • Provenance gate

    github.com

    A deterministic authorization gateway that sits between an AI agent and its tools, tracing where each parameter of a side-effect call came from, and blocking the call or asking a human when the source does not check out.

  • A markdown framework of system rules that restates Isaac Asimov's Four Laws of Robotics as a set of prioritized loss calculations for large-language-model agents to follow, published as a discovery file for agents and crawlers to read.

  • Vitruvius

    github.com

    An engineering research agent that runs a discover, read, synthesize, verify and review loop over technical questions, reading sources directly rather than inferring from titles, and publishing its own scored verifier benchmark, weaknesses included.