Zoom moves that help rather than distract

A full-frame screen recording asks the viewer to find the important thing themselves, and on a phone they will not. A zoom move points at it. Done badly, the same move makes the video unwatchable — so it is worth understanding what the automatic ones are actually based on.

It measures change, it does not track your cursor

This is the honest bit and it changes what you should expect. A browser cannot see the operating system's cursor position, so tracking clicks is not something we can do — and inventing it would mean putting a fabricated pointer in your video. Instead the analysis samples the recording, compares consecutive frames at low resolution, and finds where the picture **changed**. A change that is small and sharply localised is treated as a likely target; a change covering nearly the whole frame is read as a scroll or a page transition and ignored. It works because in a screen recording the thing that changes is almost always the thing you just did.

What gets proposed, and what does not

The analysis samples up to 900 points across the clip and counts a moment as active when more than about 0.15% of the frame changed. It discards changes that are pixel-sized noise and changes that cover three-quarters of the frame. What remains is scored — small and sharp scores higher than large and vague — and the best few are proposed, at most six and never more than roughly one per five seconds of video, with a 2.2-second exclusion around each move that is already chosen. Each proposal gets a zoom sized so the changing region fills about 45% of the frame, capped at 2.8×, with a hold of at least 1.4 seconds. They are proposals: keep the good ones, drag them, retime them, delete the rest.

Why the movement feels the way it does

The easing is smootherstep rather than the usual cubic ease-in-out, and the reason is measurable: for the same duration, cubic's peak speed is about three times its average, while smootherstep's is 1.875 times. That difference is the difference between a move that glides and one that lurches. Zoom is interpolated logarithmically, so going from 1× to 2× and from 2× to 4× take the same time and feel like the same move — linear zoom interpolation is what makes cheap zooms feel like they accelerate into the subject. The ramp is also distance-adaptive: a move that travels further across the screen is given more time rather than being whipped there.

Two rules for moves that do not annoy

First, **fewer than you think**. Two or three moves in a thirty-second demo is plenty; a move every few seconds turns the video into a fairground ride and the viewer stops reading anything. Second, **hold long enough to read**. A zoom that arrives and immediately leaves is worse than no zoom, because the viewer registers motion and gets nothing for it — that is why the minimum hold is 1.4 seconds. If two moves land close together they chain directly instead of pulling back out to wide in between, since bouncing out and back in is the most nauseating thing a demo video can do. And if a move is asked to travel too far too fast, it is flagged as squeezed rather than silently shipped.