karpathy-llm-wiki: Build a Citation-Based LLM Knowledge Wiki

Overview

karpathy-llm-wiki is a reusable Agent Skills skill for building a curated knowledge base with tools such as Claude Code, Cursor, and Codex. It gives an agent a workflow for saving source material, compiling it into Markdown wiki pages, answering questions with citations, and checking the wiki for consistency.

The project suits people who want knowledge to accumulate across research sessions: new sources can update existing pages and cross-references instead of requiring the same synthesis each time. It is a workflow and skill specification, not a hosted wiki application or a general-purpose retrieval service.

Key Features

• Ingests web pages, papers, PDFs, Markdown, text files, and pasted material into a raw/ source directory.
• Compiles sources into durable Markdown articles in a wiki/ directory.
• Answers questions using wiki pages and citations.
• Lints wiki indexes, links, and related consistency issues, with some automatic fixes.
• Maintains a global index and append-only operation log as part of the suggested structure.
• Supports Agent Skills-compatible tools, with installation or setup instructions for Claude Code, Cursor, Codex CLI, and others.
• Uses ordinary file and text search rather than requiring vector or graph search tooling.

Use Cases

• Researchers can ingest papers and articles over time, then ask a coding agent for a cited synthesis of a topic.
• Developers can maintain project or technology notes that gain cross-links as new documentation and sources are added.
• Individuals building a personal knowledge base can delegate page updates and index maintenance to an Agent Skills-compatible assistant.
• Teams or solo users can check a Markdown knowledge base for broken links and missing index entries during maintenance.

Getting Started

Install the skill with:

npx add-skill Astro-Han/karpathy-llm-wiki

See the README for tool-specific setup and usage details.

Alternatives

• cq: cq stores and validates reusable coding lessons, rather than building a cited, cross-linked wiki from source material.
• ai-memory: ai-memory preserves coding-agent activity and handoffs in a Markdown wiki, rather than organizing source material into cited reference knowledge.
• book-to-skill: book-to-skill turns documents into agent skills for on-demand use, rather than maintaining a Markdown wiki with cross-linked answers.
• cabinet: Cabinet manages a local knowledge base alongside agents and automation, rather than focusing on cited answers from source-derived wiki content.

Considerations

• The skill depends on an Agent Skills-compatible coding agent to perform ingestion, page editing, querying, and linting. It does not provide a standalone user interface or scheduled automation.
• The wiki is maintained as Markdown files, so users should expect to manage the raw/ and wiki/ directories within their own project.
• Its approach emphasizes curated pages and text search over vector retrieval. The README positions it for a bounded, curated wiki, not broad retrieval across a large raw corpus.
• The project describes itself as an unofficial implementation of Karpathy's workflow. Its documented examples and usage figures do not guarantee results for other knowledge bases.