Summary

Astro-Han/karpathy-llm-wiki provides a reusable skill for building and maintaining "LLM Wikis" compatible with Agent Skills tools like Claude Code, Cursor, and Codex. It addresses the inefficiency of LLMs re-searching raw documents by compiling ingested sources into durable, cross-referenced markdown pages. This approach allows knowledge to compound over time, offering grounded, cited answers and improving consistency, making it ideal for users who need a structured, evolving knowledge base managed by an LLM.

How It Works

The core of the project is the "LLM Wiki" paradigm, contrasting with traditional RAG. Instead of synthesizing knowledge at query time from raw chunks, this skill compiles new sources into curated markdown pages during an "Ingest" operation. The LLM actively maintains these pages, updating summaries, strengthening cross-references, and recording contradictions. A "Query" operation then retrieves information directly from these compiled wiki pages, providing citations. A "Lint" operation ensures the integrity of the wiki's structure, links, and index. The workflow utilizes distinct directories for immutable source material (raw/) and LLM-maintained knowledge pages (wiki/).

Quick Start & Requirements

• Primary install: npx add-skill Astro-Han/karpathy-llm-wiki (for Agent Skills compatible tools).
• Prerequisites: An Agent Skills compatible tool (e.g., Claude Code, Cursor, Codex CLI, OpenCode).
• Links: Full skill specification available in SKILL.md; examples provided in the examples/ directory.

Highlighted Details

• Based on a production knowledge base actively maintained daily since April 2026, featuring 94 wiki articles and 99 source materials.
• Supports ingestion of diverse sources including web pages, papers, PDFs, markdown, text files, and pasted text.
• The skill follows the agentskills.io open standard, promoting interoperability across various AI coding assistants.
• The LLM-maintained wiki model focuses on compounding knowledge and durable cross-links, differentiating it from RAG's query-time synthesis.

Maintenance & Community

The project is actively used in a production setting, with daily maintenance reported since April 2026, indicating a degree of stability and ongoing development. No specific community channels (like Discord/Slack) or notable contributors are listed in the README.

Licensing & Compatibility

• License: MIT.
• Compatibility: Permissive for commercial use and integration with closed-source applications due to the MIT license. Fully compatible with tools supporting the Agent Skills standard.

Limitations & Caveats

This skill is optimized for the "wiki model" of knowledge compounding and curated information, rather than broad, ad-hoc retrieval across very large, unstructured corpora, which is the strength of RAG. The README does not specify any alpha/beta status or known bugs.