YouTube already knows. Every video carries a replay graph built from people dragging the scrubber backwards. Rewound reads that graph and cuts where the peaks are. Paste a link. Three clips come back.
No account, no card. Three clips, free.
Not hours uploaded, not videos submitted. A four-hour podcast costs the same as a four-minute one, because the work is finding the moment, not sitting through it.
The replay data arrives with the video's metadata, before a single frame is downloaded. So it knows where to cut before it starts.
Smooth it, then throw away the opening spike. Every video's first seconds are its most replayed and they are never the interesting part.
Highest non-overlapping moments, weighted backwards toward the setup that precedes each payoff. In and out points snap to sentence boundaries.
Reframed to 9:16, captions burned in word by word, source credited on screen. A title, description and hashtags come with each one.
A language model reads the transcript and estimates which passage sounds compelling. It is a guess about text, made without ever seeing how one viewer behaved.
Reads the record of what viewers actually did. Not a prediction of what should have been interesting. Evidence of what was.