Quick Start Guide
1. Set Up Your API Keys
Before you can process content, you need to provide your own API keys in Settings:
- OpenAI API Key - For content extraction and synthesis
- Anthropic API Key (optional) - Alternative to OpenAI
- YouTube API Key - For processing YouTube videos and playlists
2. Upload Content
You can submit content in multiple ways:
- YouTube Videos - Paste any YouTube video URL
- YouTube Playlists - Paste a playlist URL to process up to 10,000 videos
- Transcript Files - Upload PDF, HTML, or Markdown files
3. Track Processing
All your submitted jobs appear on the Contribute page with real-time progress updates. You can submit multiple URLs simultaneously and watch them process in parallel.
4. Review Extracted Claims
Once extraction completes, review the extracted claims, people, and concepts. You can approve, edit, or reject each item.
What the choices mean
Catalog vs Compile
When you bring in a link or a file you choose how far to take it. Catalog just remembers it — it stores a searchable bookmark and an embedding so the source can surface later when a question needs it, with no transcript, no claim extraction, and no spend. Compile does the expensive work now: it pulls the full text and runs the model to extract claims and receipts. Catalog is the default because most sources are worth keeping but not worth compiling until a live question actually calls for them.
How questions relate: sibling, presupposition, component
When you link one question to another you name how they relate. A sibling is a different framing of the same underlying question — neither depends on the other. A presupposition is something the question quietly assumes is true; settle it and the question itself changes. A component is a sub-question you must answer on the way to answering the parent. The sentence reads "this question is a ___ of that one", where "this question" is the one you picked, not the one you are on.
Stance, confidence, and bets
A live question carries your current stance and how sure you are of it, from 0 to 100. Drag the dial to set your confidence; it saves when you let go. A stance is not a bet — a bet goes further, adding resolution criteria (what would settle it) and a date. You can revise your stance and confidence at any time as new evidence lands.
Mining sources for a question
Mining searches for sources that speak to a specific question. Quick sweeps your own backlog first — fastest and cheapest, and usually enough. Deep also casts into outside waters (the web, papers, podcasts), which surfaces more candidates but takes more time and can cost more. Nothing is ingested automatically: mining fills a shelf of candidates you choose from, and the "up to ~N sources" figure is a ceiling to shop from, not a charge.
The argument ladder: altitude, thesis, argument, support
Every compiled source is laid out as an argument tree — a skeleton. At the top (altitude 0) sits the thesis, the source’s single headline claim. Beneath it (altitude 1) are the arguments, the reasons offered for the thesis. Beneath each argument (altitude 2) is its support: the receipts, verbatim timestamped quotes that back it up. A node shown in honest-gray has no receipt yet — the view never dresses an unproven claim up as solid.
The Catch: spine, challenge, tension, bearing
After a source is compiled it raises things for you to weigh — this is the Catch. The spine is the source’s one headline claim, which you can agree with, dispute, or promote into a constellation (a thinker or idea you track in your sky). A challenge surfaces something the source assumed without arguing it. A tension flags where the source cuts against a position you’ve already taken. A bearing points out that the source speaks to one of your open questions, so you can attach it as evidence there.
Bets and predictions
A stance is what you currently believe about a question and how sure you are. A bet goes further: it adds resolution criteria — what will count as it coming true — and a date by which it resolves. A bet can back a specific position (one candidate answer to the question) or the question as a whole. Bets are how a loose opinion becomes a falsifiable prediction you can be held to.
The question lifecycle: candidate, watching, live, settled
Questions move through stages. A candidate is free and unlimited — just posed, no commitment. Watching keeps one on your radar without a stance. Making a question live commits you to actively hold it, with a stance and a spot in your live portfolio (the target range is small on purpose). Settling records your conclusion and closes it — and you’re encouraged to name what would reopen it, so a conclusion stays honest rather than final. Nothing is ever deleted; archived and settled questions can be reopened.
Key Features
YouTube Processing
- Playlist Support - Automatically expands playlists up to 10,000 videos
- Deduplication - Skips videos already in your database
- Batch Processing - Processes ~200 videos/minute
- Quota Management - Tracks YouTube API usage (10,000 units/day)
- Auto-Retry - Automatically retries failed videos
AI-Powered Extraction
- Claims - Factual assertions and arguments
- People - Speakers and individuals mentioned
- Jargon - Technical terms and definitions
- Concepts - Frameworks, heuristics, and reusable thinking tools
- Questions - Important questions raised
Prediction Markets
Use Mana (virtual currency) to bet on the truth of claims. Build reputation by correctly predicting outcomes.
Collaborative Review
Review and validate extractions from the community. Your feedback helps train the system to better match your preferences.
Troubleshooting
Job Stuck in "Pending"?
Click the "Retry" button on the job card. The system automatically detects stuck jobs and retries them, but you can manually trigger a retry if needed.
Extraction Failed?
Common causes:
- Invalid API Key - Check your OpenAI/Anthropic key in Settings
- Quota Exceeded - You've hit YouTube's daily API limit (resets at midnight PT)
- Video Unavailable - The video may be private or deleted
- No Transcript - Video doesn't have captions/subtitles
Slow Processing?
Large playlists (1,000+ videos) can take 30-60 minutes to process. The system processes videos in batches of 50 for optimal performance.