LLM Wiki (Karpathy Pattern)

A self-maintaining personal knowledge base powered by LLMs, based on Andrej Karpathy's LLM Wiki pattern.

Instead of re-searching raw documents on every question (like RAG), the LLM reads your sources once and builds a persistent, interlinked wiki that compounds over time. The more sources you feed it, the richer and more connected it gets.

Prerequisites

• Cursor (or any LLM-powered editor that reads a schema file)
• Obsidian (free) for browsing the wiki in real time

Getting Started

1. Clone the repo

git clone https://github.com/balukosuri/llm-wiki-karpathy.git
cd llm-wiki-karpathy

2. Open the project in Cursor

Cursor reads CLAUDE.md automatically and understands the wiki's structure, page formats, and workflows.

If you use a different AI agent (Claude Code, Codex, etc.), paste the contents of CLAUDE.md into your agent's context.

3. Open the same folder in Obsidian

Open the project directory as an Obsidian vault. You'll have two windows side by side — Cursor on the left where you talk to the AI, Obsidian on the right where you browse the wiki as pages appear.

4. Drop a source into raw/

Any document works:

• Product specs, design docs, or PRDs
• Meeting transcripts
• Web articles (use Obsidian Web Clipper to save pages as markdown)
• Style guides
• PDFs, reports, email threads saved as text
• Competitor documentation

5. Say "ingest"

Type this in Cursor:

ingest raw/my-document.pdf

The AI will:

1. Read the document
2. Discuss key takeaways with you
3. Create a source summary page in wiki/sources/
4. Create new pages for any products, features, personas, or concepts it finds
5. Update the glossary with new terms
6. Update the index with all new pages
7. Update the overview if the big picture shifted
8. Log everything in wiki/log.md

A single source can touch 5-15 wiki pages. Watch them appear in Obsidian in real time.

6. Ask questions

What are the main risks identified across all my sources?

The AI reads the wiki, synthesizes an answer with citations, and asks: "Should I save this as a wiki page?" If you say yes, the answer becomes a permanent analysis page. Your questions make the knowledge base richer over time.

7. Lint the wiki

Every 10 ingests or so, run a health check:

lint the wiki

The AI checks for contradictions between pages, stale claims, orphan pages with no links, missing cross-references, and inconsistent terminology. It reports what it found and asks which fixes to apply.

Repo Structure

llm-wiki-karpathy/
├── CLAUDE.md          # Schema — the AI's operating manual
├── llm-wiki.md        # Karpathy's original idea document
├── article.md         # Walkthrough article explaining this project
│
├── raw/               # Your source documents (AI reads, never writes)
│   └── .gitkeep
│
├── wiki/              # AI-generated knowledge base (AI owns this layer)
│   ├── index.md       # Master catalog — the AI reads this first on every query
│   ├── overview.md    # Big-picture synthesis (evolves with each ingest)
│   ├── glossary.md    # Terms, definitions, and style conventions
│   └── log.md         # Chronological record of all activity
│
└── .obsidian/         # Pre-configured Obsidian vault settings

How the layers work

Wiki page types

The AI creates different page types depending on what it finds in your sources:

Customizing for Your Domain

The schema file CLAUDE.md is not set in stone. Edit it to fit your needs:

• Add new page types. If your domain needs "API endpoints" or "customer segments" or "recipe variations", add them to the entity types table and tell the AI.
• Change the ingest workflow. If you want the AI to skip the discussion step and just process silently, update the workflow section.
• Adjust output formats. The AI can produce markdown pages, comparison tables, doc outlines, release notes drafts, or persona briefs. Add formats that make sense for your work.

Tips

Ingest one source at a time. You can batch-ingest, but you lose the chance to guide the AI. Stay involved — read the summaries, tell it what to emphasize, ask follow-ups during ingestion.

Save your best questions. When you ask something and get a useful answer, tell the AI to save it as an analysis page. Your explorations compound in the wiki instead of disappearing into chat history.

Use graph view often. Press Cmd+G in Obsidian. The visual map shows which pages are hubs, which are orphans, and how everything connects.

Check the glossary before writing. Open wiki/glossary.md before you write anything. It has the right terms, the wrong terms, and the reasons behind each choice.

Don't write wiki pages yourself. Your job is to find good sources and ask good questions. The AI handles the summarizing, cross-referencing, filing, and bookkeeping.

Use Cases

• Technical writers — Ingest specs, transcripts, and competitor docs. Get a living glossary, persona pages, and structured outlines without writing them yourself.
• Researchers — Feed it papers, articles, and reports over weeks. End up with a wiki that has an evolving thesis and all the connections already made.
• Product managers — Ingest PRDs, customer interviews, competitive analyses, and retros. The wiki maintains the big picture.
• Students — Ingest textbook chapters one at a time. The AI builds concept pages, links them together, and flags connections between chapters.
• Anyone accumulating knowledge — Trip planning, hobby research, health tracking, course notes, book clubs. Anything where information comes from multiple sources and you want it organized.

Credits

• Pattern by Andrej Karpathy
• Implementation and article by Balu Kosuri