Cut rate by platform, in seconds

There is a happy accident sitting at the centre of AI short-form video: the cut rate the platforms reward is roughly the clip length at which generative models stop falling apart. Cut every two seconds because TikTok likes it, and you have also never asked a model for more than two seconds of coherent motion. The format's rhythm hides the model's limit.

That accident is worth being deliberate about, because it only holds if you pick the right number for the right platform and then keep two other things in step with it.

The per-platform numbers

These are working ranges rather than rules, and they are the ones practitioners converge on for content that has something to say — tutorials, explainers, product walkthroughs, talking-to-camera. Pure entertainment and dance formats run faster; long-form essays run slower.

Platform	Cut every	Reads as
TikTok	1.5–3s	Native. Slower feels like a repost from somewhere else.
Instagram Reels	2.5–4s	Native. Faster starts to feel frantic against a more designed feed.
YouTube Shorts	3–5s	Native for tutorial content. Shorts tolerates a shot you actually have to look at.

Two things follow immediately.

A cross-posted file cannot be native on all three. If you cut once at 2 seconds and ship the same file everywhere, it is right on TikTok, brisk on Reels and twitchy on Shorts. That is usually an acceptable trade — one file at TikTok pace works everywhere — but it is a trade. If the video matters, the Shorts version wants shots held longer, which means either different edit points or a couple of extra clips generated at greater length. Aspect ratio and safe areas are a separate per-platform decision that has nothing to do with cadence; do not let one stand in for the other.

The hook is exempt. The opening 0–3 seconds is a different job from the body. It needs motion, tension or a stated problem immediately, because the decision to keep watching gets made somewhere in the first one and a half to three seconds. Then a promise lands around 3–5 seconds — what the viewer is going to get — and only after that does the body cut rate apply. Cutting at TikTok pace through the hook is fine; cutting at TikTok pace instead of having a hook is the more common mistake. Hooks in the first three seconds is the longer version of that argument.

Why the numbers match a model's stable window

Generative video degrades with duration, and it does so in a roughly predictable order. Somewhere past a handful of seconds, motion physics start to wobble. Faces drift. Chain enough clips and a character's identity has visibly moved between the first shot and the last. Duration ceilings themselves vary widely across the catalog, and plenty of models accept well over ten seconds — but accepting a length and holding together across it are different things, and the gap between them widens the longer you ask for.

So 3–5 seconds is not an arbitrary comfort zone. It is where temporal consistency holds without special handling, and it happens to sit exactly inside the cut-rate band all three platforms want. The practical consequences:

• Generate at the length you will cut at. A five-second clip trimmed to two seconds wastes three seconds of generation. A two-second clip that needed to be four means a re-roll. Decide the cut rate before you decide the shot list, and the shot list becomes a list of durations.
• Short clips also sound better. Fewer events to schedule in the audio pass means sharper foley and cleaner transients. The same constraint pays off twice.
• Shorts is where the ceiling bites. A 5-second held shot is at the top of the reliable band. If a Shorts edit wants a 7-second shot, that is a deliberate decision to spend more on stability — a first-and-last-frame approach, or two clips cross-dissolved — rather than something you get by typing a bigger number. Going meaningfully past the band also changes which models are in play, since duration ceilings differ widely.

Model choice tracks the platform too, because aspect ratio, duration ceiling and audio capability all vary: TikTok, Reels and Shorts each have a different practical shortlist.

The three axes that must agree

Cut rate on its own is meaningless. It only works when it agrees with two other things, and almost every "the pacing feels off" note is really one of these three being out of step.

Axis 1 — Cut rate. How often the picture changes.

Axis 2 — Information density. How often a new idea arrives. One idea per shot is the default; one idea across three shots is a slow reveal; three ideas in one shot is a wall.

Axis 3 — Audio rhythm. Where the beats, phrase ends and sentence breaks fall.

The two failure modes are symmetrical and both are common:

• Fast cuts over slow information reads as anxious. Nothing new is being said, but the picture keeps changing, so the viewer feels hurried without being given anything. This is the signature failure of AI short-form specifically, because clips are cheap and the temptation is to fill the timeline with them.
• Slow cuts over dense information reads as sluggish, and worse, it loses people mid-idea because there is no visual event marking that a new point has started.

The fix is always to move one axis to meet the others, not to adjust all three. In practice the audio is the one to fix against, because the ear is less forgiving than the eye: find the phrase ends in the voiceover or the beats in the track, put the cuts there, and then check whether each shot has exactly one idea in it. Anything with two gets split; anything with none gets removed.

The pattern interrupt is the exception that proves the rule. Every 10–15 seconds, break the established rhythm on purpose — a sudden hold, a jump in scale, an open loop, a change of location. The interrupt only works because the three axes were in agreement up to that point. Break a rhythm that never existed and nothing happens.

Building it on the timeline

The order that avoids rework:

1. Pick the platform and its number. Write it at the top of the shot list. Every subsequent decision reads from it.
2. Write the audio first, or find the track first. Whichever carries the rhythm. Mark the phrase ends or the beats.
3. Turn the marks into a shot list of durations. Every shot gets a length before it gets a description. Anything longer than five seconds gets flagged as a stability risk now, not after it comes back wrong.
4. Generate at those lengths, one idea per shot.
5. Assemble and check the joins. Merging videos puts the sequence together; the thing to watch for is whether the picture cut lands where the audio phrase does.
6. Time captions to the edit beat, not to the words. Auto-transcription breaks lines where the speech recogniser found gaps, which is rarely where the edit is. Short lines, landing on the same frame as the cut, read as authored; lines that straddle a cut read as an afterthought. Adding captions handles the transcription; the timing is a manual decision.

One technical note that catches people at step 5: Versely timelines default to 25 fps rather than 24, and a source clip generated at some other rate has to be conformed to the timeline. A cut rate specified in seconds is unaffected by this, but a cut rate you have counted in frames is not. Work in seconds.

FAQ

Should I cut faster if retention is dropping?

Only if the drop coincides with a shot that has nothing happening in it. A falling retention curve at a specific timestamp usually means the information stopped, not that the picture stayed still too long — and adding cuts on top of an idea that has already finished makes it worse, because now the video is both empty and busy. Find the moment the idea ran out and either cut the section or add a new idea there.

Does the cut rate apply to the hook?

The hook has its own job: motion, tension or a stated problem inside the first three seconds, then a promise by about five. Cut it however serves that, including a single held shot if the shot is arresting. The body cut rate starts once the promise has landed.

What if my clips are longer than the cut rate I need?

Trim them, but treat that as a signal for the next batch. Consistently generating five-second clips for a two-second cut rate means paying for three seconds of unused motion every shot. Generating at the cut length is the cheaper habit, and it is also the safer one, since shorter generations hold together better.

Can one edit really serve all three platforms?

One file usually can, cut at the faster end. One optimised edit cannot — the Shorts version wants longer holds and slightly denser information per shot, and that is a different assembly, not a different export. Decide up front whether the video is worth two assemblies. Most are not; the ones with a real budget behind them are.