How it works
Every answer comes from the show's own words. Nothing is invented, and nothing is imported from outside the transcripts.
1. Ingestion
When a channel is added, we fetch the publicly available caption tracks for its episodes — the same captions YouTube shows to any viewer. Those captions are cleaned, de-duplicated, split into short timestamped passages, and indexed for search. Videos without usable captions, and members-only or private videos, are skipped.
2. Retrieval
When you ask a question, it is first turned into a set of transcript searches — including the words people actually say rather than the words you typed. If your question names a guest or a specific episode, the search narrows to that episode. The most relevant passages are pulled together with their episode titles, dates and timestamps.
3. Answering
Those passages are handed to Claude, which answers from them and nothing else. It is asked to reason like an attentive listener: if a guest spends an hour warning about something, "what are they worried about?" gets a direct answer, even if nobody said those exact words. Claims are attributed to the show and its speakers rather than stated as fact, and each answer ends with the episodes and timestamps it drew on.
What it can't do
- Answer about episodes that haven't been ingested — each channel has a coverage window, and the answer will tell you when your question falls outside it.
- Tell you things nobody said on the show. It won't fill gaps with outside knowledge.
- Perfectly transcribe names. Auto-captions routinely mangle unusual names and jargon, so treat exact spellings with suspicion and check the cited timestamp.