ugc-video-prompts

UGC-Style Video Prompts

Core principle

Video models are decent at motion but worse than image models at solving "doesn't look AI" from zero — skin, lighting, and identity drift across many frames in ways a single still doesn't. Don't text-to-video a UGC clip from scratch. Generate the anchor still first (with the ugc-photo-prompts skill, or reuse one already in the media library), then animate it. The still already solved realism; the video model only has to solve movement and speech, which is what it's actually good at.

The pipeline

1. Anchor still. Already exists in the media library, or generate one with ugc-photo-prompts. This locks subject identity, wardrobe, product, lighting, and skin texture before any motion is added.
2. Motion prompt. Describe only what changes from that frame — camera behavior, body motion, speech. Don't redescribe the subject/setting; the still already carries that.
3. Model pick. Based on whether the beat needs synced dialogue, multiple combined references, or just clean simple motion — see the table below.
4. One clip = one beat. A full UGC ad is a sequence of short jump-cut beats (hook, demo, reaction, CTA), not one continuous shot. Generate each beat as its own clip off the same anchor (or a small consistent set of anchors), then cut them together on the timeline.
5. Lock the voice across beats. Same problem as appearance drift, different mechanism — describe the voice explicitly in every beat's prompt (or pin it with a referenceAudioMediaRefs clip on Seedance) so the speaker doesn't sound like a different person from one cut to the next.

Motion-prompt formula

The still already did slots 1–4 and 6–8 from the photo formula. This is just what's added for motion:

1. Seed reference. State plainly that this animates the provided still — most models infer this from startFrameMediaRef/referenceImageMediaRefs alone, but naming it keeps the prompt focused on change, not redescription.
2. Camera behavior. Default to handheld micro-shake or a static phone propped on a surface — not a smooth gimbal move or a dolly/crane. "Smooth cinematic camera move" is as strong an AI tell in video as studio lighting is in stills.
3. Subject motion. Small, continuous, natural movement — a weight shift, a blink, hair moving, a hand gesture mid-sentence. Avoid big choreographed actions unless the beat specifically calls for one (e.g., the demo beat).
4. Speech, if any. Put the exact line in quotes, written the way someone actually talks — contractions, a small pause, not ad copy. "okay wait this actually smells amazing" reads as real; "Introducing our revolutionary new formula" doesn't, regardless of delivery. If this speaker appears in more than one beat, describe the voice itself — and always include all four: pitch, tone, accent, and pace, not just one or two. "A warm, slightly raspy voice" alone is incomplete; "a warm, friendly tone, medium-pitched, neutral American accent, relaxed conversational pace" is the actual bar. Reuse that exact wording verbatim in every beat's prompt — each generate_video call has no memory of any other, so the same anchor face will get a different voice per clip unless all four are pinned down explicitly every time. Seedance can do better than words: pass the rendered audio from beat one as a referenceAudioMediaRefs input on later beats to lock the actual voice rather than approximate it in text.
5. Duration matched to the beat. Pick the shortest duration the model offers that fits one beat (often 4–6s) rather than padding a single clip to cover multiple actions — multi-action clips are where motion starts looking choreographed.
6. Negative motion cues. Add: no smooth gimbal stabilization, no cinematic camera move, no professional production polish, no overly choreographed motion, no studio lighting. Carry over the standing negative list from ugc-photo-prompts for the visual frame itself.