A personal knowledge base designed to be maintained by an LLM assistant (Claude, etc.) with human oversight. Built on a three-layer architecture: immutable raw sources, a living wiki, and a schema that teaches the LLM how to operate.
Inspired by Andrej Karpathy's LLM wiki concept — the idea that LLMs can maintain a structured, interconnected knowledge base on your behalf if you give them the right conventions.
You talk to your LLM. When you have something worth capturing — a meeting transcript, an article, a quick note — you tell it to ingest. The LLM:
- Saves the raw source (immutable — never modified)
- Extracts key information
- Creates or updates wiki pages
- Cross-links related pages with
[[wikilinks]] - Updates the index and log
When you have a question, you query. The LLM searches the wiki, synthesizes an answer, and optionally files new knowledge it discovers along the way.
Periodically, you lint. The LLM audits the wiki for contradictions, stale claims, orphan pages, and gaps worth investigating.
- Clone this repo
- Open the
CLAUDE.mdfile and customize it for your domain (replace the placeholder topics in the wiki structure) - Open the vault in Obsidian (optional but recommended — gives you a graph view of all your
[[wikilinks]]) - Start a conversation with your LLM assistant from this directory
- Tell it to ingest something: paste a transcript, link an article, drop a file in
raw/
.
├── CLAUDE.md # The schema — teaches the LLM how to operate
├── index.md # Master catalog of all wiki pages
├── log.md # Chronological record of all operations
├── raw/ # Layer 1: Immutable source documents
│ ├── articles/ # Saved articles, web pages
│ ├── assets/ # Images, PDFs, diagrams
│ ├── exports/ # Data exports, CSVs, API dumps
│ └── notes/ # Quick captures, voice-to-text, transcripts
└── wiki/ # Layer 2: LLM-maintained knowledge pages
├── concepts/ # Ideas, frameworks, mental models
├── people/ # People pages
├── personal/ # Personal notes, goals, reflections
└── projects/ # Project-specific knowledge
| Operation | What it does | When to use it |
|---|---|---|
| Ingest | Process a source into the wiki | New transcript, article, document, data dump |
| Query | Answer a question from the wiki | "What did we decide about X?" / "Summarize Y" |
| Lint | Health check the wiki | Weekly or when things feel stale |
This vault is fully compatible with Obsidian. The .obsidian/ folder contains minimal defaults with wikilinks and backlinks enabled. Open this folder as a vault in Obsidian to get:
- Graph view of all page connections
- Backlink panel (see what links to the current page)
- Full-text search across all pages
- Tag browser
- Start messy. Just paste transcripts and tell the LLM to ingest. The wiki improves over time.
- Raw sources are sacred. Never edit anything in
raw/. The wiki is the living layer. - Wikilinks are the connective tissue. The more pages link to each other, the more useful the wiki becomes.
- Customize the wiki folders. The default structure (projects/people/concepts/personal) is a starting point. Add whatever categories fit your domain.
- Lint regularly. A weekly lint catches contradictions and gaps before they compound.
- Works with any LLM. The
CLAUDE.mdschema works with Claude, GPT, or any LLM that can read files and follow instructions. Optimized for Claude Code.
MIT — do whatever you want with it.