An Obsidian vault (a self-contained folder that Obsidian treats as a workspace — this repo root) structured around Andrej Karpathy's LLM Wiki pattern: instead of querying raw documents every time (RAG), an LLM ingests sources once and continuously maintains a structured wiki — so knowledge accumulates rather than disappearing into chat history.
Inspiration:
This vault uses a three-layer architecture:
Raw Sources → LLM (Claude) → wiki/
(anywhere via CLAUDE.md entities/ concepts/ sources/ queries/
in vault)
| Layer | Path | Ownership |
|---|---|---|
| Raw sources | Anywhere in the vault (except wiki/) |
Human — organize files however you like |
| Structured wiki | wiki/ |
LLM — full write access |
| Schema & rules | .claude/CLAUDE.md |
Human — defines how the LLM should maintain the wiki |
Sources can live in Clippings/, Papers/, Notes/, or any folder you prefer anywhere in this directory — the LLM will read them wherever they are. The only rule: don't put your raw notes inside wiki/, which is LLM-managed territory.
The LLM's job: summarize, cross-reference, detect conflicts, maintain consistency across pages.
The human's job: choose what to ingest, ask questions, judge what matters.
Point Claude at any file in the vault, or paste content directly into the chat:
# Ingest a saved article
/wiki-ingest Papers/attention-is-all-you-need.md
# Paste content directly (Claude will save it to Clippings/ first, then ingest)
/wiki-ingest
<paste article text>
Claude will: summarize the source → create a wiki/sources/ page → update related entity/concept pages → update wiki/index.md.
Why URLs aren't supported: This vault uses a closed-source policy — Claude only reads from files already in the vault and never fetches from the web. This keeps
/wiki-queryanswers grounded in your own curated knowledge, prevents hallucination from stale web content, and avoids consuming extra tokens on search. Use Obsidian Web Clipper to save web pages directly intoClippings/before ingesting.
When you paste content directly, Claude saves the full text to Clippings/<date>-<slug>.md before processing — so both the original article and the wiki summary are preserved in the vault.
What is the difference between RAG and fine-tuning?
Does the vault have anything on transformer attention mechanisms?
Summarize what I know about Andrej Karpathy's work.
Answers come only from your vault — no external web search. If the vault doesn't have enough, Claude tells you what's missing.
/wiki-lint
Reports: orphan pages with no inbound links, dead wikilinks, sources in the vault that haven't been ingested yet, and pages that haven't been updated in 60+ days.
Each ingest updates wiki/index.md (navigation) and appends to wiki/log.md (audit trail).
| Tool | Version | Install |
|---|---|---|
| Obsidian | 1.12.7+ | Download from obsidian.md |
| Claude Code | latest | See Claude Code docs |
@tobilu/qmd |
2.1.0+ | npm install -g @tobilu/qmd |
qmd provides hybrid semantic search (BM25 + vector) used by /wiki-query.
obsidian CLI (bundled with Obsidian) is used for backlinks, tags, and vault-aware search.
# 1. Clone the repo
git clone <your-repo-url>
# 2. Install qmd
npm install -g @tobilu/qmd
# 3. Open the vault in Obsidian
# File → Open Vault → select the cloned folder
# 4. Index the vault with qmd (run from vault root)
qmd collection add . --name wiki
# 5. Open Claude Code in the vault directory
claudeThe .claude/CLAUDE.md schema is already included — Claude will follow the wiki maintenance rules automatically.
This repo tracks architecture only — the schema, wiki structure, and config:
✅ Tracked wiki/ structured knowledge pages
.claude/ CLAUDE.md schema + skills
.obsidian/ plugin & theme config (not workspace state)
Clippings/ directory structure only (.gitkeep)
.gitignore README.md
❌ Not tracked Clippings/** article files inside Clippings/ (personal)
Attachments/ media files
<your-folders> any other source folders you create
Clone this to get the architecture. Bring your own sources.