karpathy-llm-wiki

KO | EN

gitlite — search

#self-hosted #automation #swift #list #nlp #javascript #java #postgresql #git #claude

karpathy-llm-wiki

(★ 51)

A step-by-step guide to building a local knowledge base using Andrej Karpathy's LLM wiki method, Claude Code, and Obsidian. No vector databases, no complex infrastructure. Just markdown files that compound over time.

File Explorer

ZIP Download

README.md

# Use via CDN

jsDelivr

jsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.

Version

File

Link

Example

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

A step-by-step guide to building a local knowledge base using Andrej Karpathy's LLM wiki method, Claude Code, and Obsidian. No vector databases, no complex infrastructure. Just markdown files that compound over time.

Based on Karpathy's original idea: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

[![Watch the full build video](https://img.youtube.com/vi/lnsExa1UbnM/maxresdefault.jpg)](https://youtu.be/lnsExa1UbnM)

▶️ [Watch the full tutorial on YouTube](https://youtu.be/lnsExa1UbnM)

---

## What you need

- [Obsidian](https://obsidian.md/) (free)
- [Obsidian Web Clipper](https://obsidian.md/clipper) (Chrome extension)
- Claude Code

---

## Setup

**Step 1: Create your vault**

Download and install Obsidian. Create a new vault and inside it create a folder called `raw`. This is where all your source material will live before Claude Code processes it.

**Step 2: Install Obsidian Web Clipper**

Install the Web Clipper Chrome extension. Go to its settings and change the default save location from `clippings` to `raw`. This lets you clip any article or webpage directly into your vault with one click.

**Step 3: Open your vault in Claude Code**

Open the vault folder in your terminal and run Claude Code inside it.

**Step 4: Paste Karpathy's gist + this prompt**

Copy the full contents of [Karpathy's gist](https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f) into Claude Code, then append this prompt:

```
You are my LLM Wiki agent. Build my complete second brain from this file. Set up CLAUDE.md, index.md, log.md, and folder structure, then guide me through the first ingest step by step.
```

Claude Code will create your folder structure, `CLAUDE.md` schema, `index.md`, and `log.md` automatically.

**Step 5: Clip your first source**

Go to any article or webpage you want to save. Click the Web Clipper extension and add it to Obsidian. It will land in your `raw` folder.

**Step 6: Ingest it**

Tell Claude Code:

```
I just added a new file to raw. Please ingest it.
```

Claude Code will read the file, create wiki pages, build relationships, and update the index. Depending on the length of the source, this takes between 5 and 15 minutes.

**Step 7: Explore the graph**

Open Obsidian's graph view to see your knowledge nodes and the relationships between them. Every new source you add will connect to the existing structure automatically.

---

## Ongoing use

Every time you have a new article, transcript, or document:

1. Clip it or drop it into the `raw` folder
2. Tell Claude Code to ingest it
3. The wiki updates and links everything together

---

## Useful commands

Run a health check to find gaps and unlinked entities:
```
Run a health check on the wiki
```

Run a lint to find inconsistent data:
```
Run a lint on the wiki
```

Query your knowledge base:
```
What do I know about [topic]?
```

---

## Features

- **Auto relationships**: Claude Code links concepts across all your sources automatically
- **Health checks**: finds unlinked entities and gaps in your knowledge base
- **Lint**: flags inconsistent or missing data
- **Hot cache**: saves recent context so Claude does not re-read the entire wiki on every query
- **Plug into anything**: point any Claude project, Cursor, or other CLI tool at the wiki folder

---

## Scaling

Works well up to a few hundred documents. For millions of documents use a proper RAG pipeline with embeddings and a vector database instead.

---

## Credit

Original concept by [Andrej Karpathy](https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f).

## Author

Built by **Julian Oczkowski** — I build AI tools for knowledge work.

- 🎥 **[YouTube · @aiforwork_app](https://www.youtube.com/@aiforwork_app)** — walkthroughs and AI-for-work tutorials
- ✍️ **[Medium](https://medium.com/@julian.oczkowski)** — deep dives on product and AI workflows
- 💼 **[LinkedIn](https://www.linkedin.com/in/julianoczkowski/)** — connect and follow along

karpathy-llm-wiki

Owner Repos

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

karpathy-llm-wiki

Owner Repos

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

Download README