Fifteen seconds sounds generous until you write to it. A spoken line runs at roughly three words a second in English, less in Japanese, more in Spanish. Add a beat to breathe at the start and a beat to land the close, and the whole clip holds around forty words of usable speech. Every script the platform writes is built against that budget, and every clip is checked against it before anyone sees it.
The four beats
The delivery structure we hold to is deliberately plain:
| 0–3s | 3–7s | 7–12s | 12–15s |
|---|---|---|---|
| Earn the stop | Show the product | Demonstrate the value | Close cleanly |
The first three seconds decide whether the rest of the clip is watched at all, so they carry a situation rather than a summary: the moment before something goes wrong, the outcome someone wants, the detail people keep asking about. The middle two beats do the work the product page cannot — one uninterrupted look at the object, then one thing it visibly does. The close is short on purpose. A clip that spends its last two seconds listing features has already lost the viewer it earned.
That table is a reference, not a metronome. A tutorial script often needs the product on screen by the second second; a scenario script may hold the problem for four. What stays fixed is the total, and that the four jobs all get done.
Why the budget is enforced, not suggested
It is tempting to let a good script run long and fix it in the edit. In practice that produces one of two bad outcomes: speech sped up until it stops sounding like a person, or a line cut mid-sentence at the fifteen-second mark. Both read as synthetic, which is exactly what a UGC clip cannot afford.
So the platform checks the spoken length before it renders, and again after. If a script the system wrote runs long, it gets rewritten. If a script you supplied runs long, it is not rewritten — you are told which line pushes it over and asked to trim. A locked script stays locked; that is the whole point of supplying one.
The budget is counted differently per language. Latin-script languages are measured in characters against a wider allowance; Chinese, Japanese, and Korean against a much tighter one, because each character carries more speech. This is also why a straight translation of an English script usually fails the check: the same meaning in Japanese does not occupy the same seconds. Localisation rewrites the line for the time available rather than trimming a translation until it fits.
Where the captions go
Captions are authored from the beats, not transcribed back from the audio. That keeps them exactly in step with the voiceover and means the subtitle file exists before the video finishes rendering.
Two constraints shape them. Lines stay short enough to sit clear of the platform's own interface — the buttons down the right edge and the product card along the bottom. And nothing in the frame may imitate a tappable shopping element. A drawn-on "Buy now" button is not a design choice; it is a fake control, and it fails the check.
The three formats
The first phase ships three formats, and each is a production recipe rather than a visual theme:
- Presenter. Someone talks to camera with the product in hand. Best when the selling point is a claim someone needs to hear said out loud.
- Everyday scene. Opens on a problem, resolves it with the product in a real setting. Best when the product's value is contextual — you have to see where it goes.
- How to use. Short numbered steps ending on the result. Best when the barrier is "I don't know how this works."
In the guided workflow, the opening angle is explicit in the script before any frames are generated. If it is wrong or too generic, change it there instead of paying for a different-looking video built on the same weak idea.
What a passing clip looks like
Before a clip reaches review it has been checked for duration against the target, spoken length against the language budget, caption length against the safe area, product accuracy against the recorded facts, performer consistency against the reference frame, and expression against the target market. A clip that fails any of these is marked failed with the specific finding attached, not quietly delivered for a human to catch.
That is the trade the structure buys: the fifteen seconds are tight, but what comes out of them is predictable.