Blog · October 2, 2026

How to write an AI video ad prompt (with template)

Most AI video ads go wrong at the prompt, not the model. "A woman using our moisturiser in a bathroom, UGC style" leaves the model to decide everything: when the product appears, what she does with it, where the camera goes and what she says. The result looks fine for two seconds and then drifts.

A prompt that works reads less like a description and more like a shot list. Here is the structure we use, why each part is there, and a template you can copy.

The parts of a good AI video ad prompt

1. Reference binding. Say which attached image is which: the creator, the product, the scene. Then refer to them by those names throughout, so the model never has to guess.
2. Goal. One sentence on what the ad has to do.
3. Continuity. What must not change: the face, the outfit, the product's shape and label, the setting.
4. Timed stages. The ad split into a few stretches of time, each with the same six fields: initial state, primary action, camera, end state, transition and exact dialogue.
5. Visual style. Lighting, lens, colour, texture and how real it should look.
6. Camera and performance. Handheld or locked off, the creator's energy, when they look into the lens.
7. Audio. Voice, room sound, music or none.
8. Exclusions. What must not appear.

Why timed stages beat a paragraph

Models follow "0–3 s, 3–7 s, 7–10 s" far more reliably than a paragraph where everything happens at once. Writing an end state for every stage, and making it the next stage's initial state, keeps the action continuous instead of jumping. For a 10-second ad, three stages is usually right: a hook, the demonstration, and a closing beat with the product clearly in view.

Write the dialogue exactly

Put the spoken line in quotation marks and ask for it to be said once, as written, with nothing added. Relaxed speech runs at roughly two to three words a second, so ten seconds holds about 20 words at most. Leave room for the action too: a line of 8 to 15 words is easier to land. Never leave the dialogue to the model; it will happily invent claims you can't make.

The template

Replace everything in angle brackets. Attach the images in the order the binding lists them.

[REFERENCE BINDING]
@creator = Image 1, the character sheet.
@product = Image 2, the product sheet.
@scene = Image 3, the scene reference.

GOAL
A 10-second vertical (9:16) UGC-style ad in which @creator
<does one thing> with @product in @scene.

CONTINUITY
@creator keeps the same face, hair and outfit throughout.
@product keeps its exact shape, colour, size and label text.
The setting stays @scene.

STAGE 1 (0–3 s)
Initial state: <where everything is in the first frame>
Primary action: <the hook>
Camera: <framing and movement>
End state: <where everything is at 3 s>
Transition: <how it flows into stage 2>
Dialogue: none

STAGE 2 (3–7 s)
Initial state: <same as stage 1's end state>
Primary action: <the demonstration>
Camera: <...>
End state: <...>
Transition: <...>
Dialogue: "<the exact line>"

STAGE 3 (7–10 s)
Initial state: <...>
Primary action: <the closing beat, product clearly in view>
Camera: <...>
End state: <the last frame>
Transition: none, the ad ends
Dialogue: none

VISUAL STYLE
<lighting, lens, colour, texture, level of polish>

CAMERA AND PERFORMANCE
<handheld or steady, energy, eye contact>

AUDIO
<voice tone, room sound, music or none>

EXCLUSIONS
No on-screen text or subtitles. No logos other than the product's.
No other products. No other people.

MAINTAIN CONSISTENCY
Keep @creator, @product and @scene exactly as in their references.

A filled-in example

A fictional stainless-steel travel mug, filmed in a parked car:

STAGE 1 (0–3 s)
Initial state: @creator sits in the driver's seat, @product in the
cup holder, morning light through the windscreen.
Primary action: she lifts @product out and turns it so the logo
faces the lens.
Camera: handheld selfie angle at arm's length, slight sway.
End state: @product held at chest height, logo to camera.
Transition: continuous, same angle.
Dialogue: none

STAGE 2 (3–7 s)
Initial state: @product held at chest height, logo to camera.
Primary action: she flips the lid open with her thumb, sips, and
smiles into the lens.
Camera: same selfie angle, slow push in.
End state: lid open, @product lowered beside her face.
Transition: continuous.
Dialogue: "Still hot two hours later. I'm never going back."

Only say "still hot two hours later" if your product really does that. The line is short enough to land within the stage, and the action never asks the model to invent a part of the mug it hasn't seen.

Common mistakes

• Too much for ten seconds. Three actions and two locations become a blur. Pick one.
• Telling instead of showing. Describing the product in words when a clear reference image would hold it steadier.
• Generic style words. "Cinematic, 4K, high quality" says little. "Soft window light from the left, 26 mm phone lens, natural skin texture" says a lot.
• No exclusions. Without them, models add on-screen text that comes out garbled, extra people, or a second product.
• Letting the model write the line. Approve the words yourself, every time.

Let the studio write the plan

fairytale video builds prompts in exactly this shape. You describe the ad in a sentence or two; for each video the studio writes the concept and the timed plan, binds your creator, product and scene references to the images it attaches, and only sends it once every stage is complete. You can open any video to read the concept, the timed plan and the exact prompt and references it was made from. See AI UGC ads for the whole workflow and turning a product photo into a video ad for getting the product right.