Most AI influencer videos fail in the same way. Someone writes a paragraph that tells the whole story ("she's annoyed her group chat ignored her, so she sends a carrier pigeon, then her phone blows up"), sends it to a video model with one picture, and gets back twelve seconds of a person sitting still while something vaguely happens.
The models aren't the problem. Video models today can produce cinematic, convincing clips. They just need to be directed, not told a story. This guide shows how professionals write video prompts, using a real test we ran: the same idea rendered with our old paragraph prompt and with a proper shot list.
Direct the video shot by shot, the way an editor would cut it.
The idea: a story is a list of shots
Real short videos aren't one long take. A good 10-second reel has five to ten cuts: a close-up, a reaction, a wide shot, a detail, a payoff. Each shot does one thing, and the edit tells the story.
So write the video the way a director would hand it to an editor:
A header: how long the video is, how many cuts, and what must stay the same (the character, the place, the time of day).
One line per cut: the timestamp, the shot size, what happens, and how the camera moves.
A good AI UGC ad is decided before anything is generated: in about 60 words of script. Here is the five-beat structure our AI influencers' producer writes to, how many words fit in 20 to 30 seconds, and the mistakes that make a script sound like an ad.
Most good AI video today is image-to-video: you give the model a starting frame and a prompt. That frame does a lot of work. It fixes the face, the outfit, the room and the light, so the prompt doesn't need to describe them again.
This changes what the prompt is for. As several Kling guides put it, the prompt describes what changes, not what's there. Alibaba's official Wan prompt guide says the same: for image-to-video, the formula is simply motion plus camera movement, because the image already supplies the subject, scene and style.
Spend your effort on the frame first. Draw it with your character's reference images, check it, and only then animate it. A bad first frame can't be saved by a good video prompt. If you draw several frames for one video, plan where the camera is for each, or they all end up on the same spot: see why AI storyboard shots all look the same.
One change per shot
Each shot should contain one clear change: she looks up, the bird takes off, the phone lights up. Ask for three things in one shot and the model will do one of them, or a blurry average of all three.
Weak:
She is annoyed at her phone and decides to send a pigeon and then her phone gets lots of messages.
Better, as one shot:
Locked-off medium shot. She stares at the dark phone screen, waiting, and nothing happens. She exhales slowly and lays the phone face down on the marble table.
Name the camera move
"The camera follows her" is vague. Video models understand real camera language, and naming the move is one of the biggest upgrades you can make. The useful vocabulary:
Locked-off / fixed camera: nothing moves but the subject. Good for deadpan moments and reactions.
Push-in / pull-out: the camera moves toward or away from the subject. A push-in adds tension or focus; a pull-out reveals.
Handheld: small natural shake, like a friend filming. Reads as real life.
Tracking: the camera moves alongside the subject.
Whip-pan: a fast sideways swing with motion blur, great for carrying the viewer from one shot into the next.
Snap-zoom: a sudden zoom onto a detail, good for comic emphasis.
Rack focus: focus shifts from one thing to another in the same shot.
Shot sizes matter as much: extreme close-up, close-up, medium shot, wide shot. A reel that cuts between sizes (a wide shot, then a close-up of hands, then a reaction) feels edited and alive. A reel that stays in one medium shot feels like a webcam.
Name the move: push-in, whip-pan, snap-zoom, rack focus.
Write the cut list with timestamps
For multi-shot video, give each cut its own line and timestamp. Alibaba's Wan guide recommends exactly this format, with the shot number and a time range, for example Shot 1 [0-3 s], so the model ties each beat to the timeline. Kling's newer models accept shot lists too, either inside the prompt or as separate per-shot prompts.
Here's the cut list from our test, for a 10-second reel about a woman in 18th-century court dress who answers a silent group chat with a carrier pigeon:
10-second vertical cinematic social reel, photoreal, 8 cuts. The woman from the start frame; keep her face and look exactly in every cut. A pale Paris salon on a rainy afternoon. Cut 1 (0.0-1.4s): Medium shot, opening on the start frame: she stares at her dark phone, waiting; one eyebrow slowly rises. Slow push-in. Cut 2 (1.4-2.4s): Extreme close-up of her fingernails tapping impatiently on the marble beside the face-down phone. Locked camera. Cut 3 (2.4-3.4s): Snap-zoom to the white pigeon in the gilded cage; it tilts its head toward her. Cut 4 (3.4-4.6s): Over-the-shoulder close shot: she opens the cage door and the pigeon steps onto her hand. Cut 5 (4.6-6.0s): Wide handheld shot: she flings open the tall windows and releases the pigeon into the rain; slow-motion wingbeats. Cut 6 (6.0-7.2s): Whip-pan following the pigeon over grey rooftops, motion streaks. Cut 7 (7.2-8.6s): Extreme close-up: the phone on the marble lights up and buzzes again and again. Cut 8 (8.6-10.0s): Final lock-off medium shot: she sips her tea, smug, never looking at the phone. Sound: rain, a ticking clock, nail taps, wingbeats, phone buzzing. No dialogue. No music. No text or subtitles.
The model followed seven of the eight cuts closely: different shot sizes, the snap-zoom, the whip-pan, and the face stayed consistent throughout. Our old approach, one paragraph and a single frame, produced a static talking shot where most of the story never appeared.
Land the payoff with an end frame
The one cut that didn't land in our test was the payoff: the phone was supposed to light up, and it stayed dark. Video models are weak at changing something mid-shot, such as a screen switching on or an object appearing.
The fix is to draw the payoff yourself. Several models, including Kling, accept an end frame as well as a start frame. Draw the final moment as an image (the phone glowing with notifications, her sipping tea) and pass it as the end frame, so the video has to arrive there. This is the most reliable way to guarantee that a reveal actually appears.
Direct the sound, and keep speech for last
Modern video models generate sound with the picture, so write a sound line: rain, footsteps, a door, a phone buzzing. It makes clips feel real.
Be careful with two things:
Speech. Lip sync and timing are improving, but a line that depends on delivery (a pause, a tone, a comic beat) often comes out flat. Tell the story in pictures and keep spoken lines short, or leave them out. We cover voice in How to Make a Talking AI Influencer Video.
Music. Some models add a soundtrack even when you ask them not to. If you want a specific sound, check the result and replace the audio in editing when needed.
And never put a script or caption text in a prompt unless the model must speak it: models tend to print it on screen as subtitles.
A template you can copy
[Length]-second vertical [style] video, [number] cuts. The [character] from the start frame; keep their face and outfit exactly in every cut. [Place, time of day, weather.] Cut 1 ([start]-[end]s): [Shot size], [camera move]: [one action]. Cut 2 ([start]-[end]s): [Shot size], [camera move]: [one action]. ... Final cut ([start]-[end]s): [Shot size], lock-off: [the payoff]. Sound: [what you hear]. No dialogue. No music. No text or subtitles.
Then: start frame from your approved image, end frame from your drawn payoff, and 8 to 12 seconds total. What this costs on each model is in our reel cost tests.
A short checklist
The start frame is approved and shows the character as they should look.
Every cut has a timestamp, a shot size, a camera move and one action.
Shot sizes vary: at least one close-up and one wide shot.
The payoff is drawn as an end frame.
There's a sound line, and no script unless someone speaks.
Influgen's team of AI agents plans each post like a small production: a head writer picks the idea, a creative director draws the frames from your influencer's reference face, and the video is animated from those approved frames, with each platform's AI label on every post. You approve the storyboard before any video is made. You can see examples by niche, such as lifestyle and fashion.
Launch prices are on the pricing page. Influgen is opening in waves: create your influencer for free now, and their team starts when your spot opens.
Most AI ads give themselves away through the sound before the picture: dead air, a voice with no room around it, music fighting the words. Here are the seven fixes we now build into every talking ad, with the exact settings we use.
A sponsored reel for a small wearable gadget took us six attempts over two days. Each one failed in a different way, from an invented product to a video that stopped mid-sentence. Here is what broke, why, and what finally worked.
