After analyzing thousands of clips, the same shape keeps reappearing in the ones that perform. It isn’t a formula so much as a sequence of small promises kept to the viewer. Here’s the anatomy.
Seconds 0–3: the hook
The opening has one job — earn the next three seconds. The strongest hooks do one of three things: state a surprising claim, pose a question the viewer now needs answered, or drop into the middle of tension. What they never do is warm up. “Today I want to talk about…” is where retention goes to die.
Seconds 3–10: the promise
Once you’ve stopped the scroll, tell the viewer what they’ll get for staying. A clip that wanders loses people here. A clip that says, implicitly, “stick around and you’ll learn the one thing that fixed this” earns the rest of the runtime.
The middle: deliver, don’t pad
This is where good source material pays off. If you recorded a clear, complete thought, the middle takes care of itself. If you padded, the viewer feels it. Tight in/out points matter — trimming dead air at the edges is often the difference between a 50% and an 80% completion rate.
The end: land it
A clip that just stops feels unfinished. The best ones close the loop opened by the hook — answer the question, reveal the after, deliver the punchline. A clean ending is what turns a watch into a share.
Why detection beats intuition
You can feel all of this and still miss it in a 60-minute recording — there’s simply too much to scan. That’s the case for letting a model surface candidates: it scores every segment against this structure so the strongest moments float to the top, and you spend your judgment on curation instead of search.
The shape is learnable. Finding every instance of it in hours of footage is the part worth automating.