LLM Wiki Claude Skill — AI skill for Claude Code

llm-wiki — Claude Code Skill

Build and maintain a persistent, interlinked markdown knowledge base as a Claude Code skill. Based on [Andrej Karpathy's LLM Wiki pattern](https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f).

A single-file `SKILL.md` that teaches Claude Code how to treat a directory of markdown files (default `~/wiki/`) as a compounding, self-organizing research knowledge base. Four core operations — **ingest**, **query**, **lint**, **compile** — turn raw sources (papers, articles, URLs) into cross-linked concept pages, without requiring any external RAG service.

Why a wiki instead of RAG?

• RAG re-retrieves raw chunks on every query. Knowledge never compiles.
• LLM Wiki compiles raw sources once into cross-linked markdown pages. Contradictions are flagged. Synthesis reflects everything ingested. Every ingest makes the wiki richer — not the retrieval index.

For deep-dive research (reading papers over weeks, building a personal encyclopedia, curating a team knowledge base), the wiki pattern compounds where RAG forgets.

Features

• Four core operations: ingest sources, query the wiki, lint for health, compile (restructure) oversized content
• Divide-and-conquer pages: target 400–1200 words per page; auto-split into concepts//.md subfolders when a topic outgrows a single file
• Taxonomy discipline: every tag must come from SCHEMA.md; lint flags freeform tags
• Cross-reference enforcement: every page needs ≥2 outbound [[wikilinks]]; lint finds orphans and broken links
• Pointer files for large binaries: raw/refs/.md avoids bloating the repo with model weights, datasets, or large PDFs
• Mermaid + KaTeX: required for diagrams and formulas — ASCII art rots and is unsearchable
• Per-day log files: log/YYYYMMDD.md — grep-friendly, git-diff-friendly, naturally small
• Obsidian-compatible out of the box: point Obsidian at ~/wiki/ for graph view, backlinks,