LLM Wiki Starter — AI skill for Claude Code

llm-wiki-starter

Start a new **LLM wiki**: a knowledge base about one topic that an AI agent writes, cross-links and maintains for you.

This is Andrej Karpathy's LLM Wiki pattern, packaged. You curate sources and ask questions; Claude Code reads the sources, writes the pages, links them, and keeps the whole thing consistent. The repo holds a template, a kickoff interview that bakes your topic into the agent's rulebook, a zero-dependency web viewer and a lint pass — so every wiki you make works the same way and shares one look and feel, while each has its own page model for its own topic.

Quick start

The whole thing is operated in plain language. Open Claude Code in this repo and say what you want — for example:

**Please let's create a new LLM wiki about space ships**

That is the entire command. The agent proposes a folder (`../space-ships-wiki`), confirms it, copies the template, and then — without leaving the session — runs the kickoff interview:

• what the wiki is about, and what is explicitly out of scope
• who reads it, what they want to do with it, and the 2–5 pillars that structure the index
• which page types and entity kinds fit this topic — it proposes a model, you correct it
• an optional cross-cutting facet (era, agency, propulsion …), and whether cheat sheets fit
• how the field writes its values, what goes stale, which sources you have and what to scout

It asks in small batches, offers a default for every question so you can just say "yes", and works in your language. Then it rewrites the rulebook, `wiki.json`, README and index for your topic, writes a first-ingest plan, lints, builds the viewer, deletes itself, and hands you the page model, the first sources to ingest and the questions still open.

If you would rather script it, the shell path does the same setup:

./new-wiki.sh ../space-ships-wiki "Space Ships Wiki"   # copy the template, fill the name, git init
cd ../space-ships-wiki
claude