Things we measured while building this API, written down because we could not find them written down anywhere else. Numbers here come from real runs, and the method is named so you can repeat them.
Paste a YouTube link into ChatGPT and it answers from the description and chapter list — it never hears the video. Screenshots of what it does, which plans can add a transcript connector, the three-minute setup, and the copy-paste route for Plus and Free.
read →Three methods on the same one-hour talk: the link (ChatGPT summarises the chapter markers), the pasted transcript, and a connector. What each one gets right, and prompts that make a summary worth reading.
read →One custom connector URL on claude.ai — no OAuth — and Claude can quote, summarise and search any YouTube video with captions, on desktop and mobile too. Plus the one-line Claude Code setup and prompts that chain channel search with transcripts.
read →captions.download in the YouTube Data API returns 403 for any video you don't own, and free scrapers get IP-blocked the moment they run on a server. Why both routes fail, what self-hosting really takes, and how to get captions for any public video with one API call.
read →One LLM call for a whole transcript took 35 seconds and tripped timeouts; parallel chunks got it to 6.6, but chunks that are too small drop lines. What we measured, and how to get translated YouTube subtitles with timestamps intact by adding one parameter.
read →