UGC ad prompts for a creator who stays the same
UGC ads work because they look like a person filmed them on a phone. To get that from an AI video model, prompt for the phone look on purpose and let reference images handle who the creator is and what they hold. The prompts below cover hooks, demos and talking shots, written so you can change one line per variation and keep everything else fixed.
What makes UGC prompts work
Use two reference sets: one for the creator, one for the product. Keeping them separate means you can test the same person with a new product, or a new person with the same product, without rebuilding anything.
Ask for the phone look in plain words: selfie framing at arm's length, slight handheld movement, natural light from a window or a car windshield. Set the shape to 9:16 in the settings. Studio lighting reads as an ad and loses the point of the format.
Keep spoken lines short. People speak roughly two to three words per second, so an 8-second clip fits one sentence of 15 to 20 words. Hold the product at chest height with the front facing the lens; that is where the model keeps it most faithfully.
The prompts name the subjects: Ava and Leo are example creators and FLOW the example bottle, so write your own saved names there. Length, shape and sound are in brackets because they are settings, not prompt text.
Hook shots for the first three seconds
Prompt (6 s, 9:16, sound on): Ava sits in the driver's seat of a parked car, holding her phone at arm's length in selfie framing. She lifts FLOW into frame beside her face and says: "If you still buy plastic bottles, watch this." Daylight through the windshield, slight handheld shake.
Why it works: the car is a familiar UGC setting with flattering, even light. To test hooks, change only the spoken line and keep the rest word for word.
Prompt (6 s, 9:16): Over-the-shoulder view of Leo at a kitchen table, opening a cardboard mailer. He pulls out FLOW, turns it toward the camera and looks up at the lens with raised eyebrows. Phone camera feel, warm afternoon light, slightly cluttered counter behind.
Why it works: unboxing gives a natural reveal without dialogue. If his hands fumble the packaging, start the clip with the box already open.
Prompt (7 s, 9:16, sound on): Ava walks toward the camera on a quiet residential street, filming herself. She stops, says "Okay, I have to show you this," and turns the phone so the frame lands on FLOW in her other hand. Bright overcast daylight.
Why it works: the turn creates a pattern break in the first seconds. If the product blurs during the swing, end the clip on a steady close-up instead of mid-turn.
Demo and talking shots
Prompt (8 s, 9:16): Selfie framing of Leo in a small home gym. He shakes FLOW twice, flips the lid open and tilts it toward the lens to show the drink inside, then takes a sip. Light from a window on his left, handheld.
Why it works: each step is a simple, visible action. Keep the number of steps to three or fewer; a fourth usually gets skipped.
Prompt (8 s, 9:16, sound on): Ava sits cross-legged on a sofa, FLOW resting in her lap, label facing the camera. She looks into the lens and says: "It keeps water cold all day, and it fits in a cup holder." Soft lamp light, phone propped on a table at eye level.
Why it works: a static, propped phone suits a talking shot and keeps the face stable. The line describes the product instead of claiming weeks of personal use, which an AI creator has not had. If the mouth looks off, shorten the line before regenerating.
Running variations for ad tests
Write one master prompt and mark the slot you are testing: the setting, the opening action or the spoken line. Change one slot per variation. If you change the setting and the line at once, you cannot tell which one moved the result.
A simple test grid is three hooks times two settings, six clips in total, all with the same creator and product references. Review the six side by side for drift before spending money on traffic.
Dialogue, audio and model differences
In RefVid, turn sound on and Kling 3.0 Omni generates the voice and ambient sound with the picture. A line in quotes is said as written; English and Chinese speech sound best. Clips run 3 to 15 seconds, enough for a hook plus a short demo in one take. Describing background sound, such as light traffic outside, also helps.
On Veo 3.1, as of October 2026, according to Google's Veo docs, clips with reference images are 8 seconds, and dialogue is also written in quotes.
In RefVid the creator and the product are separate saved subjects, each with 1 to 5 photos, and up to 3 subjects can share one shot, so a test grid is the same two subjects with different scene prompts.
Check your ad platform's rules on AI-generated people and disclose where required. Use a real person's face only with their written permission, and do not present an AI creator as a real customer.
Frequently asked questions
- How long should a UGC ad line be?
- About 15 to 20 words for an 8-second clip. People speak two to three words per second, and the model needs a beat before and after the line.
- Can I use a real creator's face as a UGC reference?
- Only with that person's clear permission for this use. Many teams use a fictional character created for the brand instead, which avoids the question.
- Should an AI UGC video look low quality on purpose?
- It should look like a phone: vertical, handheld, natural light. It does not need to be blurry. Sharp phone footage reads as UGC; studio lighting does not.
- Why does the product look different between UGC hooks?
- Usually because it was described in the prompt or the references changed. Keep the product references fixed and refer to the product only by its saved name.