Second Brain Builder — Karpathy's LLM Wiki, Automated

The tool Andrej Karpathy described. Built and automated.

Turn any URL, article, meeting note, or transcript into an organized Obsidian-compatible knowledge base — with AI-generated wikilinks, entity maps, concept clusters, and synthesis pages. Zero setup. No database. No config files.

Just sources in. Organized knowledge out.

---

Why This Exists

Andrej Karpathy posted his personal LLM knowledge base workflow — a system where AI reads your sources and builds a navigable wiki of everything you've collected. The internet went wild.

This actor automates that exact system. Paste URLs or raw notes, click Start, get back a complete Second Brain structure ready to open in Obsidian or query with Claude.

---

What You Get

your-vault/
├── raw/                    ← every source preserved with metadata
├── wiki/
│   ├── sources/            ← one AI summary per source
│   ├── entities/           ← people, companies, tools, products
│   ├── concepts/           ← ideas, frameworks, methods
│   ├── synthesis/          ← cross-source comparisons & analyses
│   ├── index.md            ← master catalog of everything
│   └── log.md              ← what the AI built and why
├── outputs/                ← your Q&A and reports go here
├── CLAUDE.md               ← AI maintenance guide (add more sources later)
└── README.md               ← getting started

Open the folder as an Obsidian vault. Every page links to related pages with [[wikilinks]]. Navigate your knowledge like Wikipedia.

---

How It Works

1. Input Your Sources

Two methods — use either or both together:

URLs — paste any web pages: articles, blog posts, documentation, research papers

https://paulgraham.com/startupideas.html
https://karpathy.ai/blog/
https://docs.anthropic.com/en/docs/

Raw Content — paste text directly: meeting notes, transcripts, interview summaries, email threads

{
  "title": "Team Meeting - April 8",
  "content": "We discussed Q2 priorities..."
}

Mix and match freely in the same run.

2. AI Organizes Everything

Claude Sonnet 4.6 reads all your sources and extracts:

• Entities — people, companies, tools, products (proper nouns worth tracking)
• Concepts — ideas, frameworks, theories, methods (abstract knowledge)
• Synthesis — comparisons and analyses that span multiple sources

Every topic gets: a summary, key points, related topics with [[wikilinks]], and source citations.

3. Download Your Wiki

All files saved to Apify key-value store. Download as a folder or ZIP. Open in:

• Obsidian — browse wikilinks visually
• VS Code — code .
• Claude Code — claude . then ask anything

---

Example Output

wiki/index.md

# AI and Startups — Knowledge Base

## Entities (4)
- [[Paul Graham]] — Y Combinator founder, essay writer on startups...
- [[Andrej Karpathy]] — AI researcher, former Tesla/OpenAI...
- [[Y Combinator]] — Startup accelerator, batch model...
- [[OpenAI]] — AI research lab...

## Concepts (6)
- [[Startup Ideas]] — How to find problems worth solving...
- [[Product Market Fit]] — Achieving PMF strategies...
- [[AI Alignment]] — Ensuring AI systems behave as intended...

wiki/concepts/Startup-Ideas.md

---
tags: [startups, ideas, founder-advice]
category: concept
created: 2026-04-08
sources: 3
---

# Startup Ideas

## Summary
The best startup ideas come from problems you personally experience...

## Key Points
- Work on problems you have yourself
- Start with a narrow target market
- Ideas that seem bad often become great companies

## Related Topics
- [[Founder Mindset]]
- [[Product Market Fit]]
- [[Paul Graham]]

## Sources
- [How to Get Startup Ideas — Paul Graham](https://paulgraham.com/startupideas.html)
- [Startup School Lecture 2024](https://...)

---

Use Cases

Personal Knowledge Management Research projects, learning new skills, building expertise in any domain.

Business Intelligence Organize competitor research, industry news, market analysis into a queryable wiki.

Content Creation Build research bases for newsletters, blogs, podcasts, video scripts.

Academic Research Structure papers, sources, and notes for thesis work or literature reviews.

Meeting Intelligence Paste transcripts and notes. Get back organized action items, decisions, and concept maps.

---

Asking Questions After

Once your wiki is built, open it in Claude Code and ask:

"What are the main themes across all these sources?"
"What does source A say vs source B about [topic]?"
"Who are the key people in this knowledge base and what do they believe?"
"What gaps should I research next?"
"Write a 500-word briefing on [concept]"

The CLAUDE.md file tells your AI exactly how to maintain and extend the wiki.

---

Pricing

Pay only when a knowledge base is successfully created:

Knowledge Base Size	Topics Generated	Price
Small	1–5 topics	$0.50
Medium	6–15 topics	$1.50
Large	16–30 topics	$3.00
Huge	31+ topics	$5.00

No charges for failed runs. No hidden fees.

---

Input Options

Parameter	Type	Description
urls	array	Web pages to scrape
rawContent	array	Direct text input (title + content)
topic	string	What this knowledge base is about
maxPages	integer	Max sources to process (default: 50)
includeScreenshots	boolean	Save visual snapshots (default: false)
useLightpanda	boolean	10x faster scraping for large batches
useProfile	boolean	Persistent browser profile for auth bypass

---

Tips for Best Results

URL selection:

• Use article pages and long-form content, not homepages
• Mix multiple authors and perspectives
• 10–50 sources gives the best AI organization

Topic description:

• Good: "AI safety research and alignment strategies"
• Good: "B2B SaaS marketing and growth tactics for 2026"
• Bad: "stuff" or "everything"

Page limits:

• 10–20 pages: quick test or narrow topic
• 50–100 pages: standard knowledge base
• 200–500 pages: comprehensive research

---

FAQ

What websites can I scrape? Almost any public website. Uses agent-browser (Vercel Labs) which handles JavaScript-heavy sites, SPAs, and dynamic content. Cannot access paywalled or login-protected content unless you enable persistent profile mode.

Do I need an AI API key? No. AI is included. We handle the Vercel AI Gateway integration with Claude Sonnet 4.6.

Is this compatible with Obsidian? 100%. Proper YAML frontmatter, [[wikilinks]], and markdown structure that Obsidian loves natively.

Can I add more sources later? Yes. Drop new files into raw/ and tell Claude: "Read new files in raw/ and update wiki/ following CLAUDE.md rules."

How is this different from bookmarking? Bookmarks are unorganized lists. This gives you organized topics with summaries, connections between concepts, and a wiki that's immediately queryable by AI.

What's the difference between Lightpanda mode and normal mode? Lightpanda is a purpose-built headless browser engine — 10x faster and 10x less memory than Chrome. Use it for large batches. Note: Lightpanda doesn't support persistent profiles.

---

Credits

Inspired by Andrej Karpathy's personal LLM knowledge base system.

Built by OpenClaw AI — making AI-powered knowledge management accessible to everyone.