Skip to content

Latest commit

 

History

4 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

LLM Wiki — Second Brain Scaffold

License: MIT Contributing Code of Conduct Built with Claude Code Works with Obsidian

A workshop scaffold for building a personal knowledge base maintained by an LLM agent, inspired by Andrej Karpathy's vision of a personal wiki.

The core idea: instead of querying raw documents at every turn (RAG), the LLM incrementally builds and maintains a persistent wiki — a structured, interlinked collection of Markdown pages that compounds over time. Cross-references, contradictions, and synthesis are computed once and kept current, not re-derived on every question.

Table of Contents

What you need

How it works

raw/              ← you drop sources here (articles, notes, transcripts, PDFs)
wiki/             ← the LLM writes and maintains all pages here
  └── index.md   ← auto-generated table of contents
log.md            ← append-only operation history
CLAUDE.md         ← the schema: naming conventions, page format, linking rules
.claude/commands/ ← the four slash commands

You curate sources and ask questions. The LLM does all the bookkeeping — summarizing, cross-referencing, flagging contradictions, keeping the index current.

Open wiki/ in Obsidian to browse the knowledge graph as the LLM builds it.

The four commands

Command What it does
/ingest Read new sources in raw/, compile wiki pages, weave backlinks, flag contradictions, move processed files to raw/processed/
/query <question> Answer from the wiki with citations; flag anything sourced from general knowledge
/save <content> Turn a conversation answer or any text into a new wiki page
/lint Health-check: broken links, orphans, contradictions, concepts missing their own page

Getting started

  1. Clone or copy this repo as your wiki folder.
  2. Open it in Claude Code (or any agent that reads CLAUDE.md).
  3. Drop a source file into raw/ and run /ingest.
  4. Open wiki/ in Obsidian to explore the result.

Edit CLAUDE.md to adapt the schema to your domain — page types, frontmatter fields, linking conventions. The schema is intentionally minimal; co-evolve it with your agent as conventions settle.

Workshop

This scaffold is used in the AI Club LLM Wiki workshop. The session covers:

  • The difference between RAG and a compounding wiki
  • How CLAUDE.md acts as a contract between you and the agent
  • Live demo: ingesting sources and querying the result in Obsidian
  • Designing your own schema for a domain of your choice

Contributing

Contributions are welcome — see CONTRIBUTING.md for how to propose changes, the areas that are good candidates for a PR, and the style this project follows.

Code of Conduct

This project follows the Contributor Covenant Code of Conduct. By participating, you're expected to uphold it.

License

Released under the MIT License.

Credits

Concept inspired by Andrej Karpathy's writing on personal knowledge management and the broader idea of LLMs as tireless wiki maintainers.

About

LLM-maintained second brain: Claude Code ingests your notes into a persistent, cross-linked Obsidian wiki instead of re-querying raw docs.

Topics

Resources

Code of conduct

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors