# How to prompt AI video tools for a product launch video

A product launch video prompt guide built from real launch videos and each tool's own docs: shot lists, prompt formulas, references, text and sound.

Published 2026-11-11. https://motionreference.com/blog/ai-video-prompts-product-launch

You can't prompt a whole product launch video into existence. You can prompt its shots. This guide shows how to turn a real launch video's structure into a shot list, write a prompt for each shot in the way each tool's own documentation recommends, and keep text, logos and interface out of the parts AI still gets wrong.

Everything about the tools comes from their own guides, checked in November 2026 and linked throughout. The example shots come from launch videos in the Motion Reference library.

## What the tools can do today

A single generation is a short clip, so a launch video is built from several. Here's what each tool's documentation says, as of November 2026:

| Tool | Clip length per generation | Aspect ratios | Sound generated | What the vendor says about on-screen text |
| --- | --- | --- | --- | --- |
| [Google Veo 3.1](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/veo/3-1-generate) | 4, 6 or 8 s | 16:9, 9:16 | Yes | Nothing |
| [Google Gemini Omni 1.2 Flash](https://aistudio.google.com/docs/omni) | 3–15 s | 16:9, 9:16 | Yes | Renders text "correct and readable" |
| [Kling 3.0](https://kling.ai/blog/kling-ai-prompt-guide) | Up to 15 s | Check Kling's docs | Yes | "Precise lettering"; keep text short |
| [Kling 4.0](https://kling.ai/feature/kling-4-ai-video-generation) | 3–30 s | 21:9, 16:9, 1:1, 9:16 | Not stated | Not stated |
| [Luma Ray3.2](https://lumalabs.ai/ray3-2) | 5 or 10 s | Check Luma's docs | No, add it in the edit | "Never ask AI to render readable text or logos" |
| [Runway Gen-4.5](https://academy.runwayml.com/prompt-guide) | Check Runway's docs | Check Runway's docs | Not stated | Small labels come back "as gibberish text" |

## Start from a reference, not a blank prompt

It's tempting to write one prompt for the whole video. Start instead with a launch video whose structure fits yours, and copy its beats.

Take MiniMax's launch for its H3 image model. It's a transformation: one jellyfish image becomes prints, couture, origami, lamps, magazines and game worlds, at a pace that keeps building.

[MiniMax H3 Image Launch by MiniMax](https://motionreference.com/reference/78b33562-ffdf-4abf-ab5c-3c262b6268ac) (Markdown: https://motionreference.com/reference/78b33562-ffdf-4abf-ab5c-3c262b6268ac.md)
Its five beats, from the library's analysis:

| Time | Beat | What happens |
| --- | --- | --- |
| 0:00–0:09 | Iterate the concept | The photo becomes line art and prints, placed in outdoor and city settings |
| 0:09–0:16 | Fashion | The motif becomes couture and digital billboards |
| 0:16–0:28 | Objects | Origami diagrams, a pendant lamp, a clothing line |
| 0:28–0:43 | Media | Magazines, posters, data screens, game worlds |
| 0:43–0:50 | Reveal | Top-down: the creator on a floor covered in the work, then the end card |

That's your shot list. Each beat becomes two or three shots of 4 to 8 seconds, which fits what every tool above can generate in one go.

## The prompt formula the tools agree on

The vendors use different words for the same structure:

- **Google (Veo 3.1):** [Cinematography] + [Subject] + [Action] + [Context] + [Style & Ambiance] ([Google Cloud](https://cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-veo-3-1))
- **Luma:** Subject + Action + Setting + Camera + Lighting + Style ([Luma](https://lumalabs.ai/news/write-ai-video-prompts))
- **Kling 3.0:** subject, action, scene, camera, lighting and mood, style ([Kling](https://kling.ai/blog/kling-ai-prompt-guide))
- **Runway (text to video):** "[Camera] shot of [a subject/object] [action] in [environment]" ([Runway Academy](https://academy.runwayml.com/prompt-guide))

So every shot prompt answers six questions: what's on screen, what it does, where, how the camera sees it, how it's lit, and what it should look like. Luma adds a practical rule: put the most important details in the first 20 to 30 words.

## Write the shots

Here's the jellyfish beat sheet rewritten for a hypothetical product launch: a ceramic travel mug with a wave-pattern glaze. The motif (the wave) plays the jellyfish's role. These are example prompts in the vendors' formats, not tested outputs.

**Beat 1, the motif (image to video).** Generate and approve a still first, then prompt only the motion. Luma says image-to-video "inherits the lighting, color, and composition" from your still, and Runway's image-to-video formula is "The camera [motion description] as the subject [action]".

> The camera slowly pushes in as steam rises from the travel mug and the wave glaze catches the morning light.

**Beat 2, the motif in the world (text to video, Veo-style order).**

> Medium shot, slow tracking left. A commuter on a rain-wet city platform lifts a ceramic travel mug with a blue wave-pattern glaze. Soft overcast light, reflections on the tiles. Muted blues and greys, quiet documentary style.

**Beat 3, the motif as objects (one move per shot).**

> Close-up, static camera. On a pale oak table, the same blue wave pattern appears printed on a notebook, a tote bag and a ceramic plate beside the travel mug. Even daylight from a window, minimal and editorial.

**Beat 4, the hero product shot (Luma's product template).** Luma's own example for products is a "Slow 360-degree orbit around [PRODUCT NAME] on [SURFACE]":

> Slow 360-degree orbit around a ceramic travel mug with a blue wave-pattern glaze on a wet slate surface. Soft studio light, shallow depth of field, premium product style.

**Beat 5, the reveal.** Generate the top-down wide shot, then add the product name, line and logo in the edit. Don't put them in the prompt (more on that below).

> Top-down wide shot, static camera. A designer sits on a studio floor surrounded by prints, mugs and fabric swatches, all carrying the same blue wave pattern. Bright, even light, clean editorial style.

## Rules that hold across tools

- **One camera move per shot.** Kling: "one main move per shot is usually better" ([Kling](https://kling.ai/blog/kling-ai-camera-control-video-guide)). Luma: "Use only ONE camera movement per clip to avoid unpredictable results" ([Luma](https://lumalabs.ai/news/write-ai-video-prompts)). Split complex moves into separate shots and cut them together.
- **Lock the frame, then animate.** Runway's [product video guide](https://dev.runway.com/learn/generate-product-videos): "Write the script first, then generate the key frames, then generate video from those frames." The approved still decides composition, product and palette, so the video model only has to decide motion.
- **Show the product from several angles.** Runway: "One hero shot is not enough", and "Whatever the reference does not show, the model invents." Keep the product references identical across shots, and change only the scene.
- **Say what you want, not what you don't.** Runway: "Describe what you want to see rather than what you don't want" ([Runway Academy](https://academy.runwayml.com/prompt-guide)). Google suggests listing unwanted elements rather than writing "no" ([Google prompt guide](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/video/video-gen-prompt-guide)).
- **Generate each aspect ratio directly.** Runway: "Generate to each ratio rather than cropping a landscape master", because cropping "throws away the composition you approved."

## Keep text, logos and interface out of the prompt

This is where the vendors disagree. Google says Gemini Omni renders text "correct and readable", and Kling claims "precise lettering" for 3.0. Luma says "Never ask AI to render readable text or logos". Runway's product guide says text on the product "should come from your reference images rather than being generated".

Independent research supports caution. An October 2026 benchmark tested 11 video models on scenes with text, including advertisements, and found that even the best got about one word in four wrong ([VTR-Bench](https://arxiv.org/abs/2610.01499)).

Our advice for a launch video:

- **Titles, claims and logos:** add them in the edit. Launch videos rely on on-screen type: kinetic typography appears in 49 of the 66 product launch videos in the library.
- **Text on the product:** supply it through reference images of the real product.
- **Your interface:** screen-capture the real product and animate it in the edit. None of the vendors offers guidance on generating product UI, and a launch video promises what people will actually see.

## Multi-shot prompts and sound

Some tools can generate several shots in one go:

- **Veo 3.1:** timestamp prompting assigns a shot to each segment, for example `[00:00-00:02] Medium shot … [00:02-00:04] Reverse shot …` ([Google Cloud](https://cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-veo-3-1)).
- **Gemini Omni:** timing in natural language ("After 3 seconds…") or as timecodes ("[0-5s] …") ([Google](https://aistudio.google.com/docs/omni)).
- **Kling 3.0:** Multi-Shot, automatic or with each shot set by hand ([Kling](https://kling.ai/blog/kling-ai-prompt-guide)).

For the rest, generate shot by shot and cut in an editor. Runway suggests using the last frame of one clip as the start of the next.

For sound, Veo 3.1 takes dialogue in quotation marks, sound effects as `SFX: …` and background sound as `Ambient noise: …`. Luma's Ray3.2 doesn't generate sound, so add music and effects in the edit. Most launch videos are music-led anyway: 43 of the 66 product launches in the library are, so many teams score the edit rather than generate sound shot by shot.

## Using an agent to plan the shots

The slow part is the shot list, not the prompts. If you use an AI agent such as Claude or Codex, connect it to Motion Reference over [MCP](/connect) and ask it to find a launch video with the structure you want, read its timed moments, and turn them into a shot list with a prompt per shot in your tool's format. You review the list, generate the stills, then animate.

## FAQ

### Can AI make a whole product launch video?

Not in one prompt. Single generations run from about 4 to 30 seconds depending on the tool, and vendors advise against relying on generated text and logos. Use AI for shots, and assemble, title and score them in an editor.

### Which AI video tool should I use for a product video?

It depends on the shot. Check clip length, aspect ratios and sound against your shot list using the table above. If you need native sound, Veo 3.1, Gemini Omni and Kling 3.0 generate it. For product consistency, look for reference-image support: Veo 3.1 takes up to three references of a product, and Kling 4.0 up to 10 images.

### How long should a product launch video prompt be?

Luma suggests 100 to 150 words, with the key details in the first 20 to 30. Runway warns that "extremely complex, multi-paragraph prompts" leave the model less room and can produce unnatural results. One shot per prompt keeps it short.

---

Find a launch video with the structure you need in the [Motion Reference library](/), and use its timed moments as your shot list.