How It Works

Three Layers

Every vault follows Karpathy’s three-layer architecture:

Layer	Contents	Who Owns It
Raw sources	Articles, PDFs, videos, tweets, gists in raw/	Immutable — LLM reads but never modifies
The wiki	LLM-generated markdown in wiki/ — source-notes, entities, concepts, comparisons	LLM owns entirely
The schema	CLAUDE.md — conventions, domain tags, page structure rules	Human defines, LLM follows

The raw layer is the source of truth. The wiki layer is the LLM’s compiled understanding. The schema layer governs how the wiki is structured.

Four Operations

Operation	Command	What It Does
Ingest	/ingest <source>	Process a source into wiki pages (3-7 pages per source)
Query	/query "question"	Ask questions, get cited answers, --save to compound
Lint	/lint	Health-check wiki (8 checks, --fix auto-repair)
Promote	/promote <vault>	Graduate reusable knowledge between vaults

These four operations form a cycle. Ingest brings knowledge in. Query explores and synthesizes. Lint maintains quality. Promote distributes reusable learnings across vaults.

One Engine, Many Vaults

The engine (this repo with 25 Claude Code skills) operates on separate vault directories. Each vault is its own git repo.

llm-wiki/                     # The engine (public, shared)
├── .claude/skills/           # 25 skills
├── vaults/                   # Gitignored -- your private data
│   ├── meta/                 # Long-lived cross-project knowledge
│   ├── project-alpha/        # Per-project, archived when done
│   └── personal/             # Private vault

Vaults are data, not applications. No skills, no logic — just markdown, raw sources, and a thin CLAUDE.md for conventions. The engine operates on them.

Each vault is its own git repo. This gives you independent lifecycles, different privacy levels, clean history, and size independence. Karpathy himself uses separate wikis per research area.

The Compounding Effect

The system compounds in three ways:

1. Ingest compounds — each new source enriches existing entity and concept pages
2. Query compounds — saved answers (--save) become first-class wiki pages
3. Promote compounds — reusable knowledge from finished projects enriches the meta vault

This is the core insight: the wiki is a persistent, growing artifact — not re-derived each query like RAG.