How to write a script for AI video that survives generation

A screenplay assumes a crew that can hold one idea for three minutes. An actor crosses a room, changes his mind, picks up the phone, and the camera follows all of it in a single take. A video engine holds one idea for about five seconds and then stops. Scripts written for the first medium fall apart in the second — not because the writing is bad, but because every line contains three shots and the generator has to guess which one you meant. Here is how to write a script that comes out of the machine looking like the thing you wrote.

Write in beats the length of a clip

The unit of an AI film is not the scene, it is the clip, and the clip length is decided by the engine before you type a word:

• Motion One · Wan — 5 or 10 s, available on every account
• Seedance 1.0 Lite and Seedance 1.0 Pro — 3 to 12 s
• Hailuo 02 — 6 or 10 s
• Kling 2.5 Turbo — 5 or 10 s
• Veo 3 Fast — a fixed 8 s
• Seedance 2.5 — 4 to 30 s

Your plan caps it a second time: 10 s on the free tier, 15 s on Hobby, 20 s on Creator, 30 s on Studio and above. So the practical beat is five to ten seconds, and the test is embarrassingly simple. Say what happens in the beat out loud. If it takes you longer than ten seconds to say it, it is two beats.

A beat is one subject, one action, one camera position. That is the whole specification.

One action per shot

Here is a line that reads well and renders badly:

NADIA pushes through the door, scans the room, and sits down across from the man in the grey coat.

Three actions and two subjects in eighteen words. An engine asked to do all of it at once produces an average of the three — a woman vaguely near a door, vaguely sitting, the man arriving from nowhere. Split it and every piece becomes shootable:

1. Nadia pushes through a glass door into a dark lobby and stops. (Tracking, 5 s)
2. Her eyes move across the room, reading it. (Close-up, 3 to 5 s)
3. She sits down across from a man in a grey coat. (Wide, 5 s)

Three clips, fifteen seconds, one legible idea each. The cut does the work the sentence was trying to do. This is also why a storyboard is the right shape for the script — you are not writing paragraphs, you are writing a numbered list of shots that happen to have prose attached.

Where the words go

Nothing in eroq syncs a mouth to a recorded line. There is no lip-sync, which means a script built on close-ups of people delivering dialogue is a script betting on a feature that does not exist. Three things do work:

• Voice-over. Write the words as narration, record them in the Voice studio, and lay the MP3 over the picture in your editor. Most AI shorts are narrated for this reason, and narration is a legitimate form, not a consolation prize.
• Engine sound. Veo 3 Fast is the only engine here that renders a soundtrack with the clip — ambience, effects, dialogue. You do not direct it the way you direct a recorded line, but a room that needs to sound like a room gets one.
• Text you add later. eroq does not draw titles, captions or logos. Those go on in Premiere or After Effects, and you should plan the empty part of the frame for them while you are still writing.

Write speech in the script anyway. You need to know what a character means before you can shoot her face knowing it.

Draft the prose in the Text studio, then shoot in the Video studio

The Text studio is the cheapest room in the building — RP+ at 3 credits a completion, RP mini at 1 — and it is where a scene should be argued with before it costs a render. Use the "scene" register for immersive prose, "messaging" for short texting-style exchanges.

The output formatting does a favor you will only appreciate on the second film: *actions and thoughts* render dimmed and italic while **emphasis** renders bold, so speech and action separate at a glance. When you go to convert a scene into shots, the italic lines are your picture and the plain lines are your sound. The list writes itself.

Convert each beat into one paragraph

A prompt that renders is one flowing paragraph of thirty to eighty words naming the subject, the motion, the camera, the light and the mood — in that order, because that is roughly the order of importance to the model. A complete beat, ready to paste:

A woman in a grey wool coat pushes through a glass office door into a dark lobby after hours, stops dead, and lets the door swing shut behind her; tracking shot from behind at shoulder height, the only light a bank of vending machines glowing to her left, rain running down the glass wall beyond, an uneasy and quiet mood.

Seventy words, one action, one camera, one light source. The composer offers twenty-four camera moves — Static, Slow pan, Push-in, Tracking, Orbit, Handheld, Close-up, Wide, Pull-back, Dolly zoom, Crane up, Crane down, Whip pan, FPV drone, Aerial pull-back, POV, Snorricam, Bullet time, Slider, Roll, Dutch angle, Overhead, Crash zoom, Steadicam — and written directions inside the prompt ("wide shot", "dolly", "tracking") are detected too, so the two ways of saying it agree rather than fight.

If your beat is still three words and a feeling, press Enhance. The prompt enhancer rewrites a rough note into a director-grade paragraph and costs nothing, so there is no reason to hand the engine your first draft.

A worked six-beat page

A ninety-second opening, written as six beats and shot at 5 s on Seedance 1.0 Lite at 60 credits each:

1. Establishing — the office tower from across the street, rain. Wide, static.
2. Nadia through the glass door, stopping. Tracking.
3. Her eyes reading the room. Close-up.
4. The man in the grey coat, not looking up. Wide.
5. She sits. Slow pan.
6. The folder slides across the table. Close-up, push-in.

Six renders, 360 credits, one take each. Two of those shots will disappoint you — they always do — and re-rendering both costs another 120. Budget a first pass, not a masterpiece, and you will spend roughly a third of what you spend discovering the same notes at four takes a scene.

Write the beats. Time them out loud. Shoot them one at a time.

FAQ

How long should one beat be?

Five to ten seconds of screen time, which is one subject doing one thing. Engines run from 3 s to 30 s, and your plan caps the maximum at 10, 15, 20 or 30 s.

Can I write dialogue at all?

Write it, then deliver it as voice-over — there is no lip-sync. Veo 3 Fast renders its own audio including spoken sound, but you do not control the performance the way you control a recorded track.

Do I have to write the prompts myself?

No. Enhance rewrites a rough line into a full director-grade prompt for free, and Surprise me invents one when you have a look but no scene yet.

Got a page of beats? Open the video studio and shoot beat one.