2026-10-25

How to make AI UGC ads in Google Flow

A free, honest guide to making UGC-style video ads in Google Flow: what it costs, the workflow, two full prompts, and how to disclose AI people.

You make an AI UGC ad in Google Flow by writing one detailed prompt that describes a believable person, a real-looking room, what their hands do and the exact line they say to camera. Flow then generates a short vertical clip with the voice built in. The parts that make it hold up are casting a consistent face, keeping the prompt strict about hands and lip sync, and adding very little in the edit.

Google Flow and Veo are Google products. PromptVault is not affiliated with Google.

Can you make UGC ads with AI?

Yes. A UGC-style ad is a person talking to their phone, not a polished pitch. That format suits video models well: one person, one room, natural light, a few sentences.

What AI can't do is make that person a real customer. The person in the clip does not exist. So you can use them as a presenter, but you should never present them as a genuine buyer giving a testimonial. More on that in the disclosure section below.

What you need and what it costs

You need three paid tools at most, plus a free chatbot:

None of these needs its top plan for this workflow.

On cost per video, the course page gives these figures for buying generation raw through a model API:

• A ten-second segment runs about 36 cents on the cheapest engine.
• The same segment runs about $1.34 on the most expensive engine.
• A thirty-second ad made by chaining segments is roughly $1.10 of compute.

Through a platform like Flow, you pay for a plan instead, so your real cost depends on that plan. Check the plan's current pricing before you start.

The workflow, step by step

This is the order the course follows. You can do the short version (steps 2 and 4 only) with the free prompts below.

1. Cast the face. Start from reference photos, collected on Pinterest, so the person doesn't look like the default face a model always makes. Use photos you have the right to use, and don't recreate a real, identifiable person without their permission.
2. Write the prompt. Give two reference photos to Claude and have it write the prompt as text. It writes the prompt; it doesn't make the image.
3. Get a second opinion and plan the room. Run the same brief past ChatGPT. Use it to write a prompt for an empty-room plate too: an image of the set with nobody in it, so the room stays fixed across clips.
4. Generate in Google Flow. Use Flow's Characters to keep one person, and its scene timeline to line clips up. Flow also has an agent that can walk the chain of clips for you.
5. Go longer if you need to. For clips around twenty seconds, the course uses Seedance 2.5 (through Riverflow), because it holds a face across that length. Riverflow also gives you multi-reference and an edit mode that locks the face and changes everything around it.
6. Swap in your voice (optional). Clone your own voice in ElevenLabs v3. The course uses two audio effects so the voice sounds like it's in the generated room.
7. Fix the lip sync. Use HeyGen to redraw the mouth so it matches your recording. The course drives HeyGen's API from Cursor so this is a repeatable step.
8. Finish in CapCut. Trim and export. Mostly, this step is about what not to add.

The two free sample prompts

These are the two recipes the course page shares for free. Both use the same settings:

• Tool: Google Flow
• Model setting shown on the page: Omni 1.1 Flash
• Length: 10 seconds
• Aspect ratio: 9:16
• Resolution: 720p

Each one also comes with a reference frame, which you can find on the course page. The last sentence of each prompt tells the model to use that image only for style, light and camera feel.

The spoken lines are demo lines about AI itself. For a real ad, change the line to something that is true of your product.

Laundry fold

Raw vertical 9:16 smartphone video that looks exactly like a real person's unedited social post, not an ad: natural available light, true-to-life skin with pores, fine flyaway hairs and small imperfections, no beauty filter, natural blinks and micro-expressions, precise lip sync, relaxed unhurried hand movement, gentle handheld sway, the soft background blur of a phone lens. One person only, anatomically correct hands with five fingers, the same face and clothes in every frame, no morphing, no warped objects. Clean phone-mic audio with light room tone, no music. A man in his late thirties with a short salt-and-pepper beard and a faded blue T-shirt, stands at a kitchen table with a neat pile of folded laundry, morning daylight from the window behind him, a mug and a radiator in the background. He lifts one already-folded white T-shirt off the pile, holds it up briefly, sets it back down and looks at the lens. Speaking straight to the lens in a casual, warm, slightly fast, unscripted tone, they say: "Everyone says AI can't do hands or fabric. I just picked up a T-shirt, put it down and talked to you. Not one second of this was filmed." No subtitles, no captions, no on-screen text, no logos or brand names visible. Use the reference image for art style, palette, lighting and camera feel only; new characters and a new scene.

What to check in the result:

• Five fingers on each hand, and the hands grip the T-shirt cleanly.
• The T-shirt keeps its shape when it's lifted and set back down.
• The beard and the blue T-shirt look the same from first frame to last.
• The mouth matches the words, with no music under the voice.
• No captions, text or logos have crept in.

Rainy bus stop

Raw vertical 9:16 smartphone video that looks exactly like a real person's unedited social post, not an ad: natural available light, true-to-life skin with pores, fine flyaway hairs and small imperfections, no beauty filter, natural blinks and micro-expressions, precise lip sync, relaxed unhurried hand movement, gentle handheld sway, the soft background blur of a phone lens. One person only, anatomically correct hands with five fingers, the same face and clothes in every frame, no morphing, no warped objects. Clean phone-mic audio with light room tone, no music. A woman in her mid twenties with a short blonde pixie cut, a small silver nose stud and a dark green raincoat with the hood down, stands under a glass bus shelter on a rainy evening street, rain streaking the glass, wet pavement, a red bus passing blurred behind her. She pushes a damp strand of hair off her forehead and smiles slightly. Speaking straight to the lens in a casual, warm, slightly fast, unscripted tone, they say: "Rain on the glass, a bus going past, my hair's a mess. This is what makes people stop scrolling. And nobody filmed this. I typed it." No subtitles, no captions, no on-screen text, no logos or brand names visible. Use the reference image for art style, palette, lighting and camera feel only; new characters and a new scene.

What to check in the result:

• The nose stud stays in place, and the hood stays down.
• The hand that moves her hair has five fingers and doesn't blend into her face.
• The bus behind her stays blurred and doesn't turn into a second person.
• Rain on the glass looks like rain, not smears on the lens.
• The voice sounds like a phone mic outdoors, with no music.

How the prompts are built

Both prompts follow the same order. First comes a fixed block about how a real phone video looks. Next, a block of rules: one person, five fingers, same face and clothes, no morphing. After that come the audio, then the person, the place and one simple action. The spoken line follows, written out in quotes. Last, the things to leave out. To make your own, keep the first two blocks as they are and change the rest.

Making the person look real and consistent

The course page names a few specific moves for this:

• Cast from real reference photos, not from a model's default face.
• Make an identity plate first. It's a straight-on portrait of your character, used as the reference image for every clip. You then add the line: Keep this person's face, hair and clothes exactly as in the reference image.
• Make an empty-room plate so the set stays the same before anyone stands in it.
• Ask for imperfections in the prompt. Pores, flyaway hairs, no beauty filter, natural blinks, gentle handheld sway.
• Use Flow's Characters to keep one person across clips.
• Pick an engine that holds a face if you need around twenty seconds in one take.

Disclosing AI people in ads

Rules change, so check the platform's current policy before you publish. Here is what the official pages say as of this writing.

TikTok. TikTok's Ads Manager has a disclaimer toggle for ads with image, video or audio that is fully AI-generated or substantially altered by AI. You switch it on while creating the ad. See Add disclaimers to ads in TikTok Ads Manager and the misleading and false content ad policy. TikTok's AI-generated content guidelines also ask creators to label realistic AI content. Its Integrity and Authenticity guidelines bar using a private adult's likeness without permission.

Meta. For ads about social issues, elections or politics, Meta requires you to disclose photorealistic AI media, including a realistic-looking person who does not exist. See Meta's policy page. The same page says Meta will use automated detection to spot AI-made ad media and show an "AI info" label in "About this ad". For other kinds of ads, check the platform's current policy.

Whatever the rules say, one thing doesn't change. Don't present an AI actor as a real customer, and don't put words like "I've been using it every night" in their mouth unless you label the clip clearly as AI. Only make claims about your product that are true.

Common problems

• Extra or merged fingers. Keep the rules block in, keep the hand action simple, and generate again.
• The face drifts between clips. Use an identity plate as the reference and add the "keep this person's face" line. Or use Flow's Characters.
• Captions or text appear. Keep the "no subtitles, no captions, no on-screen text" line at the end.
• Music under the voice. Keep "Clean phone-mic audio with light room tone, no music."
• It looks like an ad. Cut back the props, the script and the edit. The format works because it looks unedited.
• The voice sounds generated. Clone your own voice in ElevenLabs, then re-sync the mouth in HeyGen.
• A 30-second take falls apart. Chain shorter segments instead of asking for one long take.

For more Google Flow prompts, each shown with the unedited clip it produced and what every part of the prompt does, see the free prompts page.

Where can I learn the full UGC workflow?

The two prompts above are free. Ultra-Realistic AI UGC covers the full workflow in ten lessons, recorded end to end. Each lesson's prompt comes as a PDF. It also includes 50 UGC prompts across eight niches and the recipes for the locked clips on the page. It's a one-time payment of $29 (or ₹1,499), with lifetime access.

The course this guide comes from

Ultra-Realistic AI UGC: Ten lessons, recorded end to end: ads that look like a person made them.