Founder Book

A self-hostable, LLM-maintained knowledge base built from how founders actually think.

Founder Book is a free, open-source tool that turns founder YouTube talks and essays into a searchable, AI-queryable knowledge base — an LLM wiki. It points itself at YouTube channels and essay sites (Garry Tan, Y Combinator, YC Root Access, Paul Graham, Sam Altman, …), pulls the transcripts and articles, runs them through Gemini to extract people, companies, and topics, and weaves it all into an interconnected Markdown wiki you can query in plain English from your terminal — with cited, grounded answers (retrieval-augmented generation).

It then keeps itself up to date on its own: a background process watches those sources and ingests new videos and essays automatically while you use the tool — never fetching the same thing twice.

Inspired by Andrej Karpathy's idea of an LLM Wiki — a personal knowledge base that a language model continuously reads, writes, and cross-links for you, instead of a static pile of notes you maintain by hand. Founder Book is one concrete take on that idea, specialized for startup/founder knowledge and made fully automatic.

---

The problem

The best thinking from founders and investors is buried in hundreds of hours of YouTube videos and dozens of long essays. When you have a specific question — how do I price for enterprise? what did they say about hiring the first engineer? how do I tell if I have product-market fit? — the answer is often a throwaway line 40 minutes into one talk and a single paragraph in someone else's essay. Getting it means scrolling, scrubbing through video after video, and skimming article after article.

You shouldn't have to watch and read everything to get one answer. You should be able to just ask — and have an LLM respond using the mental models, frameworks, and advice that were actually in those videos and essays, pull the relevant ones up for you, and, when you want to go deeper, hand you the exact mentions and source links so you can use them directly.

That is what Founder Book does. Ask a question in plain English → it retrieves the relevant mental models from across every source, synthesizes a grounded answer, and cites the specific videos/essays (with the mentions behind each claim) so the originals are one click away. No scrolling required.

What it does

YouTube channels / Essay sites / Local PDFs
        │
        ▼  discover new items (incremental, never re-fetches)
   extract transcripts (parallel Tor circuits)  +  scrape essays
        │
        ▼  Gemini analysis → entities, topics, summaries, claims, quotes
   Structured wiki (Markdown, Obsidian-compatible, fully wikilinked)
        │
        ▼
   Ask questions in natural language (RAG over the wiki + raw transcripts)

• Build a wiki of sources/ (one page per video/essay), entities/ (people, companies, products) and topics/ (fundraising, PMF, hiring, …), all cross-linked.
• Ask questions conversationally; answers are grounded in retrieved pages + raw transcripts, with citations, and refuse to invent facts.
• Stay current automatically — see Automatic updates.

---

What's included in this repo

This repository ships both the engine and a large starter corpus so it works the moment you clone it:

Source	Type	Items
Garry Tan	YouTube	~170 transcripts
Y Combinator	YouTube	~800 transcripts
YC Root Access	YouTube	~130 transcripts
Paul Graham	Essays	~215 essays
Sam Altman	Essays	~110 essays
Generated wiki	Markdown	~1,500 sources · ~3,000 entities · ~3,000 topics

All source content belongs to its original authors and is included for research/study. See Content & attribution.

Privacy: this repo intentionally contains no API keys, no personal notes/ideas, and no query history. Those live only on the maintainer's machine and are git-ignored. See .gitignore.

---

Who it's for

• Founders, operators, and aspiring entrepreneurs who want to search what top founders and investors actually said — startup advice, fundraising tactics, hiring, product-market fit, growth — instead of re-watching hours of video.
• Anyone building an AI "second brain" or personal knowledge management (PKM) system who wants a working, self-hostable example.
• Developers looking for a clean reference implementation of a retrieval-augmented generation (RAG) pipeline over YouTube transcripts and essays, with automatic, incremental ingestion.
• Researchers and writers studying startup, venture-capital, and AI discourse from primary sources.

Quick start

Fastest — one command

git clone https://github.com/ckryptickunal/Founder-Book.git
cd Founder-Book
bash install.sh

install.sh installs the Python dependencies and launches a short, guided setup that:

• walks you through your Gemini API key (with an optional live "test this key" check),
• lets you add a YouTube key (optional — only for discovering new videos),
• lets you pick a model (the default works on Google's free tier, so a brand-new key just works), and
• registers a terminal keyword so you can open the wiki from anywhere.

Then, from any terminal:

founderbook                 # opens the interactive wiki (use whatever keyword you chose)
founderbook "how do I price for enterprise?"   # one-shot answer

First run with no key? The CLI runs the same guided setup automatically — you're never dropped at a cryptic error. Re-run setup anytime with python3 setup_cli.py (or just keys: --keys, just the keyword: --launcher).

Manual setup (alternative)

pip install -r requirements.txt
brew install tor                 # macOS (Linux: apt install tor) — for transcript extraction
cp .env.example .env             # then add YOUTUBE_API_KEY / GEMINI_API_KEY
python3 query_wiki.py "What does Paul Graham say about doing great work?"

The wiki ships pre-built, so you can query immediately. To rebuild or extend it, read on.

---

Automatic updates (the headline feature)

You should never have to remember to "go fetch the new videos." Founder Book does it for you.

How it works

When you start the interactive CLI (python3 query_wiki.py), it launches auto_sync.py as a detached background process. That process:

1. Reads the watch-list in sources.json.
2. For each YouTube channel, walks the uploads feed newest-first and stops as soon as it reaches videos you already have — so a routine check costs a couple of cheap metadata calls, not a full re-crawl. Only genuinely new videos are extracted (via Tor).
3. For each essay site, reads the live index and downloads only essays not already on disk.
4. Ingests everything new into the wiki with Gemini — in one pass.

It is smart about not repeating work: an item counts as "known" if it is on disk, recorded in a channel's extract-state file, or already in the ingest manifest. State lives in wiki/automation_state.json.

It is safe and unobtrusive:

• A lockfile ensures only one sync runs at a time, even with several CLI windows open.
• A cooldown (default 6 h/source) stops it re-crawling on every launch.
• It degrades gracefully: no Tor → it still discovers + queues videos and processes essays; no API key → it fetches but defers ingestion and tells you. One failure never aborts the rest.
• It runs in its own process group writing to automation.log, so the REPL is never blocked.

Controlling it

Inside the Q&A CLI:

Command	Action
/sync	Show last-sync status per source
/sync now	Force a background sync right now

Standalone / scripted:

python3 auto_sync.py                # one pass (respects cooldown)
python3 auto_sync.py --force        # sync now, ignore cooldown
python3 auto_sync.py --dry-run      # show what WOULD be fetched; change nothing
python3 auto_sync.py --source "Garry Tan"   # one source only
python3 auto_sync.py --loop 360     # daemon: sync every 6 hours
python3 auto_sync.py --status       # print last-sync summary

Run it as a real daemon with cron/launchd — see docs/AUTOMATION.md.

Watching a new source

Add an entry to sources.json and you're done — the next sync picks it up:

{
  "youtube_channels": [
    { "name": "My Channel", "query": "https://www.youtube.com/@somehandle",
      "channel_id": null, "folder": "My Channel" }
  ],
  "essay_sources": [
    { "name": "Some Blog", "kind": "generic",
      "index_url": "https://example.com/essays", "base_url": "https://example.com",
      "folder": "Some Blog", "id_prefix": "sb" }
  ]
}

(channel_id is resolved from query once and cached back automatically.)

---

The scripts

Content acquisition

Script	What it does
auto_sync.py	Background engine. Discovers + fetches + ingests new videos/essays.
fetch_new_essays.py	Incrementally scrapes new essays from authors' websites.
extract_channel.py	Bulk transcript extraction via parallel Tor circuits.
transcriptor.py	Interactive single-video / full-channel extraction + channel resolution.
fetch_transcript.py	Fetch and print one transcript (quick caption check).
retry_failed.py	Retry videos in failed_videos.json through fresh Tor circuits.
fetch_essays.py	One-time essay import from public HuggingFace datasets.

Wiki management

Script	What it does
ingest.py	Analyze transcripts with Gemini → build/update wiki pages (parallel).
ingest_file.py	Ingest local PDFs / text / markdown / HTML.
query_wiki.py	Interactive Q&A CLI (RAG) — and launcher for the background sync.
query_skills.py	Task-specific answer skills (summary, comparison, …).
lint_wiki.py	Health check: broken links, orphans, stats.
run_pipeline.py	Full one-shot pipeline: retry → essays → ingest → verify.
background.py	Detaches auto_sync.py from the CLI; status/force helpers.

A full walkthrough of each is in docs/ARCHITECTURE.md.

---

Asking questions

python3 query_wiki.py                       # interactive REPL
python3 query_wiki.py "..."                 # one-shot

Interactive commands: /sync, /ingest <path>, /stats, /pages, /skills, /save, /quit. Answers are synthesized from retrieved wiki pages and the underlying raw transcripts, with citations. If the wiki has no evidence, the answer says so rather than guessing.

---

Project structure

.
├── sources.json              # watch-list for the auto-sync (channels + essay sites)
├── auto_sync.py              # background discovery → extract → ingest engine
├── fetch_new_essays.py       # incremental essay scraper
├── background.py             # detached launcher used by the CLI
├── extract_channel.py        # Tor-parallel transcript extraction
├── transcriptor.py           # channel resolution + metadata + transcript writer
├── ingest.py                 # Gemini wiki ingestion
├── ingest_file.py            # local file ingestion (PDF/txt/md/html)
├── query_wiki.py             # interactive Q&A CLI
├── query_skills.py           # answer skills
├── lint_wiki.py / run_pipeline.py / retry_failed.py / fetch_*.py
├── requirements.txt
├── .env.example
│
├── Garry Tan/  Y Combinator/  YC Root Access/   # raw YouTube transcripts (.txt)
├── Paul Graham/  Sam Altman/                     # essays as transcript files
│
└── wiki/                     # the generated knowledge base
    ├── sources/   entities/   topics/            # cross-linked Markdown pages
    ├── index.md                                  # catalog of all pages
    ├── ingested.json                             # ingest manifest (dedupe state)
    └── schema.md                                 # wiki format spec

---

Environment variables

Variable	Required	Used by
YOUTUBE_API_KEY	yes (for new YouTube videos)	transcriptor.py, extract_channel.py, auto_sync.py, retry_failed.py
GEMINI_API_KEY	yes (for ingestion & Q&A)	ingest.py, query_wiki.py, lint_wiki.py, ingest_file.py
GEMINI_MODEL	no	Override ingestion model (default gemini-3.1-flash-lite)
GEMINI_MODEL_QUERY	no	Override Q&A model
GEMINI_MODEL_LINT	no	Override lint model

---

Troubleshooting

• Tor not running / port 9050 refused → tor (or brew services start tor). The background sync detects this and defers extraction until Tor is up.
• YouTube rate-limit / IP block → shouldn't happen (each request uses a unique Tor exit IP); if it does, re-run — extraction resumes from where it stopped.
• Gemini errors during ingest → check GEMINI_API_KEY; lower ingest.py --workers.
• Missing video titles ("Unknown") → invalid/rate-limited YOUTUBE_API_KEY; transcripts still download, only metadata is missing.

More in docs/AUTOMATION.md and docs/ARCHITECTURE.md.

---

FAQ

What is Founder Book? Founder Book is a free, open-source, self-hostable knowledge base that turns founder YouTube talks and essays into a searchable, cross-linked Markdown wiki you can query in natural language. It's a concrete implementation of an LLM-maintained wiki (an idea popularized by Andrej Karpathy) focused on startup and founder knowledge.

How is this different from searching YouTube or Google? Instead of returning links, Founder Book reads the full transcripts, extracts the people, companies, and topics, and synthesizes a direct, cited answer from the primary source material — grounded retrieval (RAG), not guesswork.

Does it keep itself up to date? Yes. A background process (auto_sync.py) automatically discovers and ingests new videos and essays from the channels and blogs you watch, while you use the tool, without ever re-processing what it already has.

Which founders and sources are included? A starter corpus from Garry Tan, Y Combinator, YC Root Access, Paul Graham, and Sam Altman — and you can add any YouTube channel or essay site in sources.json.

What AI model does it use? Google Gemini for ingestion and question-answering (model is configurable). You bring your own API keys; nothing is sent anywhere else.

Is my data private? Yes. Your API keys, personal notes/ideas, query history, and any locally-ingested documents stay on your machine and are git-ignored — they are never published.

Can I use it with Obsidian? Yes — the wiki is plain Markdown with YAML frontmatter and [[wikilinks]], so it opens as a graph in Obsidian out of the box.

Do I need Tor? Only to extract new YouTube transcripts at scale (it avoids IP blocks). Querying the existing wiki and reading essays works without it.

---

Content & attribution

• Concept: inspired by Andrej Karpathy's idea of an LLM-maintained wiki.
• Source content: transcripts and essays are the work of their respective authors (Garry Tan, Y Combinator, Paul Graham, Sam Altman, and others) and are included here for research, study, and personal knowledge-management. All rights remain with the original creators. If you are a rights-holder and would like content removed, please open an issue.
• Code: released under the MIT License.

Contributions welcome — see CONTRIBUTING.md.

---

_{Topics: LLM wiki · AI second brain · RAG over YouTube transcripts · founder knowledge base · searchable startup advice · Y Combinator / Paul Graham / Sam Altman essays · Gemini RAG · personal knowledge management · Obsidian-compatible wiki · automatic YouTube transcript ingestion · open-source AI knowledge base.}