How long should an AI video prompt be?

As long as its shot list: one short row for each two to three seconds of video. A 15-second clip is about six rows.

Does this work on every video model?

It's written for current multi-shot models such as Seedance and Kling, which cut between cameras inside one take. Yuumeh checks each prompt against these rules and flags rows that won't work.

What does "Cut to new camera" do?

It tells the model the next row is a different shot from a different camera position, so it cuts instead of drifting.

How to prompt AI video as a shot list

The most common reason an AI video looks lifeless is the prompt: one paragraph, read by the model as one long, static shot. Write the clip as a shot list instead, with a new camera every two to three seconds. This is the rule Yuumeh's AI producer follows for every video.

The rule

One row per shot. The opening shot goes on the first row.
Start each later row with Cut to new camera:
Cut every two to three seconds. That's about six shots in a 15-second clip.
Say the shot's size and angle, then what happens in it: wide, close-up, extreme close-up, insert, low angle, high angle, side view, over the shoulder, POV.
Vary them. Two close-ups in a row read as one shot.

An example

A 15-second clip in six shots

Wide shot, dawn: a baker unlocks a small corner shop, her breath showing in the cold.
Cut to new camera: close-up, hands dusting flour across a wooden bench.
Cut to new camera: insert, the oven door swinging open on an orange glow.
Cut to new camera: low angle, trays of dough sliding in.
Cut to new camera: over the shoulder, the first customer waiting at the window.
Cut to new camera: extreme close-up, steam rising off a cut loaf. Hard cut to black.

Not every shot is story

Editors cut away to breathe. Give a clip atmosphere shots and cutaways: dust drifting through a beam of light, rain coming off a roof, a lantern swaying, someone's reaction. They set the rhythm and make the story shots land.

Timing lives in the cuts

A joke or a threat needs three shots, not one. Set it up in the first. Hold the beat in a shot of its own: everyone frozen, silence. Pay it off in the next, then show the reaction. End the clip on a button, such as an extreme close-up or a hard cut to black.

Mistakes to avoid

A paragraph. The model reads prose as one take.
The camera and three actions on one row. One row, one shot, one thing happening.
A cut every second. Past one cut every two seconds a scene turns into a trailer.
Lines of speech buried in a row. Spoken lines get rows of their own: see how to write spoken lines.

Naming what's in the shot

When a card has several pictures or characters plugged in, name them in the prompt so the model knows which is which. On Yuumeh you write @ and the input's name, for example "@Mara walks towards @Image1".

Questions

How long should an AI video prompt be?: As long as its shot list: one short row for each two to three seconds of video. A 15-second clip is about six rows.
Does this work on every video model?: It's written for current multi-shot models such as Seedance and Kling, which cut between cameras inside one take. Yuumeh checks each prompt against these rules and flags rows that won't work.
What does "Cut to new camera" do?: It tells the model the next row is a different shot from a different camera position, so it cuts instead of drifting.

Reviewed 11 November 2026