Karpathy LLM Wiki
karpathy-llm-wiki
A skill for a personal knowledge base the model maintains: ingest sources, compile cited articles, lint integrity
Install
npx add-skill Astro-Han/karpathy-llm-wikiWorks with any tool that supports the Agent Skills standard.
This is third-party code. Review the repository files before installing.
What it does
The skill turns an agent into the editor of a personal wiki following Karpathy's idea: the model maintains the knowledge base while the human chooses sources and asks questions. It works with two folders: raw for immutable sources and wiki for compiled articles with an index and an operation log. There are three operations: ingest collects a source and compiles or updates articles, query answers with citations to wiki pages, and lint checks links, the index and whether facts match the sources. A grounding invariant requires every load-bearing fact in the wiki to appear verbatim in the linked source, and a script verifies this across the whole base.
Who it is for. For researchers, analysts and technical specialists who accumulate knowledge on topics.
Good fit when
- You want a knowledge base the model maintains
- You need answers cited to your own articles rather than raw sources
- You accumulate research articles over time
Not a fit when
- You need a one-off lookup without maintaining a base
- Plain RAG over a large corpus without manual curation fits better
Example request
Ingest this article into the wiki and update the related pagesLimitations
The skill deliberately skips vector or graph search, numeric confidence scores and scheduled runs, relying on grep, reading and whole-wiki lint. Ingest needs web or file access via your tools, otherwise a source is pasted manually.
How to disable. Remove the karpathy-llm-wiki folder from your agent's skills directory.
Security check
- The skill reads and writes files in the raw and wiki folders and fetches sources from the web
README in short
The README packages Karpathy's LLM wiki idea into one installable skill. The agent gathers sources into raw, compiles durable knowledge pages into wiki, answers with citations and lints the base for integrity. It compares the approach to RAG and lays out design boundaries: vector search, numeric confidence scores and scheduled runs are deliberately left out. Installation is via npx add-skill, compatible with any Agent Skills tool. MIT licensed.
SKILL.md
--- name: karpathy-llm-wiki description: "Use when building or maintaining a personal LLM-powered knowledge base. Triggers: ingesting sources into a wiki, querying wiki knowledge, linting wiki quality, 'add to wiki', 'what do I know about', or any mention of 'LLM wiki' or 'Karpathy wiki'." --- # Karpathy LLM Wiki Build and maintain a personal knowledge base using LLMs. You manage two directories: `raw/` (immutable source material) and `wiki/` (compiled knowledge articles). Sources go into raw/, you compile them into wiki articles, and the wiki compounds over time.
FAQ
How does this differ from RAG?
Knowledge lives in curated markdown pages, and synthesis happens during ingest and maintenance rather than per query.
What sources can I ingest?
Web pages, papers, PDFs, markdown and text, all converted to markdown in raw and compiled into wiki.
Related
A CLI for every Google Workspace API with JSON output and agent skills: Drive, Gmail, Calendar, Sheets and more
Local search over Markdown notes, docs and meeting transcripts: keywords, semantic search and reranking, with an MCP server
A task manager for AI-driven development: breaks a PRD into dependent tasks and guides the agent through them via MCP or CLI
Anthropic's official plugins for knowledge work roles: sales, support, product, marketing, legal, finance, data and enterprise search