If your AI-generated product videos look beautiful but don’t sell, you’re in the majority: IAB’s 2026 video report found 96% of smaller video buyers say they’re not satisfied with how they use generative AI for ad creative. The problem isn’t the model. It’s the prompt. Well-structured prompts keep packaging intact, show one clear product action, and end on a readable frame — exactly what a scroller needs in three to eight seconds.
Why structure matters more than camera tricks
Most public prompt libraries teach camera moves and visual flair: orbit shots, dolly pushes, dramatic color. Those are useful for cinematographers, but they’re the wrong starting point when your goal is a short ad that communicates a single product benefit. A product ad has three operational jobs: preserve the SKU’s visual identity, show one buyer-relevant action, and finish on a frame that makes the product unmistakable — ideally with the label readable with or without sound.
Every line of your prompt must serve those jobs. Extra moves, vague descriptors, or repeated product descriptions give the model license to “improve” the asset — which often means redesigning packaging, swapping colors, or shortening the product’s screen time. Structure reduces drift and makes results repeatable across a campaign.
The six-slot prompt framework (use in any tool)
Build every product-video prompt from six functional slots plus an exclusion line. Learn these once and adapt them to most video models and formats.
- Format — clip length and aspect ratio (for example: 5s / 9:16 for a short vertical ad).
- Product lock — name the product and explicitly state what must not change (label, cap, color, shape, count). This prevents the model from redesigning your SKU.
- Moment — define one action, written as 2–3 beats (for example: reach, lift, set down). Keep it sequential and singular.
- Camera — allow only one camera move and a framing (e.g., slow push-in, close-up). A single purposeful move reads as direction; multiple moves read as drift.
- Look — specify a light source and three colors to use consistently across assets (for brand cohesion).
- Ending — say exactly where the camera stops and what the final frame should show (product in hand, label readable, pack shot on a neutral surface). This gives each clip a destination.
Exclude — close with 4–5 things the model typically gets wrong (example: no floating labels, no changing cap color, no background clutter). If your tool provides a negative-prompt field, move exclusions there for reliability.
Two slots do most of the work: Product lock preserves identity; Ending fixes the shot’s destination so motion reads as direction, not aimless movement.
Prompt core (paste into any model)
Format: [length]-second [9:16 / 1:1 / 4:5 / 16:9] clip. Product lock: [what the product is] + what must not change (label, color, shape, count). Moment: one action in 2 or 3 beats (reach, lift, set down). Camera: one move + framing (slow push-in, close-up). Look: light source + 3 named colors. Ending: where the camera stops and what the last frame shows. Exclude: [4 or 5 things you keep seeing go wrong].
Use this core as a template. Fill the brackets, remove anything the specific generator handles as a parameter (duration, resolution), and run a short batch to test constraints before scaling.
Image-to-video vs text-to-video: pick based on precision
- Image-to-video (I2V) — Start from a product photo when exact label, texture or shape must remain intact. The image provides identity, so the prompt can focus on motion and timing instead of re-describing the SKU.
- Text-to-video (T2V) — Use text when you need conceptual scenes: ingredient shots, problem/solution moments, or hands-only demos where the product acts as a prop. Treat T2V results as stand-ins unless you iterate with real product imagery later.
Rule of thumb: anything where packaging matters should start from an image. Concept work and storyboard exploration can begin from text.
Small adapters for different tools
Most generators accept the same functional prompt, but platforms differ on parameters. A few practical rules reduce failed runs:
- If the tool has explicit duration and resolution settings, set them there and remove length/resolution from the prompt text. Some engines ignore those phrases inside natural language.
- If the model generates audio, add a separate audio sentence after the Look line describing voiceover style or background sound.
- If your tool has a negative/‘exclude’ prompt field, move your exclusions into that field to keep the main prompt focused.
- When a platform accepts a reference image, use I2V prompts and cut redundant product description — the image supplies identity.
Test once and lock a template for the tool you use. The functional elements — subject, context, action, ending — are consistent across major providers, but parameter handling varies by vendor.
How to operationalize this in campaigns
- Start with short clips (3–8 seconds) that show one clear action and end on a readable product frame. Short form reduces production cost and speeds iteration.
- Create a small matrix: three motions (e.g., lift, pour, uncap) × two lighting looks × one product lock. That yields controlled variations you can A/B test without rewriting assets.
- Export in native aspect ratios (9:16 for Reels/TikTok, 4:5 for feed) so safe zones and cuts remain correct on delivery platforms.
- When outputs drift, tighten Product lock and Ending before adding more style language; models need constraints, not extra adjectives.
Adoption of generative video is high — IAB found nearly two in three video ad buyers use generative AI and roughly a third of their ad assets already lean on it — but dissatisfaction persists because many prompts are under-specified. Fixing prompt structure addresses the most common failure modes and makes creative runs predictable and reusable.
Next steps and what to watch
Pick one priority SKU, capture a clean product photo, and build three 5-second I2V prompts using the six-slot framework. Run them as a batch and measure attention on the final frame; if people aren’t reading the label, tighten the Ending until the product is the unmistakable focal point.
Watch two practical shifts: clearer UI controls for duration/aspect ratio (which reduces the need to state them in text) and wider support for negative-prompt fields — both make exclusions and constraints easier to enforce across runs.
