AI Video Prompts: One Structure That Works Across Every Model

The shared structure of a video prompt

Vendors order the parts slightly differently (Google suggests cinematography, subject, action, context and style for Veo; ByteDance suggests subject and action, scene, light, camera, style and constraints for Seedance), but the parts are the same. Checked in November 2026 against each model's official guide.

• Subject: who or what is in frame, with two or three stable visual traits. "A potter in her sixties with a clay-stained linen apron."
• Action: one, and specific. "Lifts a freshly glazed bowl from the wheel."
• Scene or context: where and when. "Small workshop at dusk, rain on the window."
• Camera: shot size, angle and one movement. "Medium shot at eye level, slow dolly-in."
• Light and style: type of light, colour and overall look. "Warm desk-lamp light, documentary style."
• Sound: ambience, effects and dialogue, if the model generates audio. Name the speaker.
• Restrictions: what you don't want, in whatever form the model accepts (a separate field, a constraint phrase, or nothing at all).

One clip, one shot, one action. If the story needs several beats, use the model's multi-shot format or generate several clips and cut them together.

A template and an example to copy

This template works as a base in any model. Fill in the gaps, then adapt the sound and negatives to your tool's format.

[Shot size and angle], [one camera movement]. [Subject with 2–3 visual traits] [one action] in [location, time of day]. [Lighting]. [Style and mood]. [Who speaks] says, "[short line]". [Sound effects]. [Ambient sound].

Medium shot at eye level, slow dolly-in. A young florist with a green canvas apron and rolled-up sleeves ties a bouquet of sunflowers with twine in a small shop at closing time. Soft window light from the left, warm tones. Natural documentary style, calm mood. The florist says, "Last one for today." Twine pulled tight, scissors snapping shut. A quiet street outside.

In Veo the last two sentences would become "SFX:" and "Ambient noise:"; in Seedance they'd go inside < >; in Runway Gen-4.5 you'd drop them, since its audio isn't verified. Here's how each model differs.

How the prompt changes by model

• Veo 3.1 (Google): English prompts; dialogue in quotes, "SFX:" and "Ambient noise:"; multi-shot with [00:00-00:02] timestamps; a negativePrompt field that takes a list of terms. 4, 6 or 8 s clips that can be extended. See the Veo 3 prompts guide.
• Kling 4.0 (Kuaishou): the role of each asset (@Image1, @Video1), general settings, a "0–X s" timeline and audio; dialogue in quotes with the language stated and the voice described, in 9 languages. It advises against negative prompts. See the Kling prompts guide.
• Seedance 2.0 and 2.5 (ByteDance): literal symbols, with （ ） for music, < > for effects, { } for dialogue and 【 】 for subtitles; "Shot 1 / Shot 2" with one move per shot; constraint phrases instead of a negative field. 2.5 adds "0-3s" timings and 11 dialogue languages. See the Seedance prompts guide.
• Wan (Alibaba): English or Chinese prompts, an overall description then "Shot 1 [0–3 s]"; a fixed label for each character and the action before the dialogue; a negative_prompt field. Wan 2.2 has open weights but no sound. See the Wan prompts guide.
• MiniMax Hailuo: H3 uses sections (integrated description, English soundscape and music) and dialogue as (S1) plus <d> tags, with camera moves in prose. Hailuo 02 and 2.3 use bracketed camera commands such as [Push in]. See the Hailuo prompts guide.
• Runway Gen-4.5: positive phrasing only, since negatives can produce the opposite; "locked-off" for a still camera; complex moves described phase by phase. See the Runway prompts guide.

Camera moves: the glossary

Every model understands plain English film vocabulary: "close-up", "low-angle shot", "tracking shot", "slow dolly-in". The AI video camera glossary explains every shot size, angle and move, with the exact form each model documents (Hailuo's brackets, Runway's "locked-off", Veo's arc shot and whip pan).

Low-angle tracking shot following a courier on a red bicycle through a narrow cobbled street at dawn, the camera moving alongside at walking pace. Cool morning light, long shadows, realistic style.

Picking a model: the comparison

If you haven't settled on a tool, the AI video generator comparison lays out, from official documentation only, each model's modes, clip length, audio, prompt language, negatives and access, and rewrites one prompt in every model's syntax. If you used Sora, which has been retired, the Sora alternatives guide shows how to port your prompts.

Prompts by use case

Each kind of video comes with its own decisions. These PromptNow templates ask for what matters in a form and hand back a prompt ready to paste:

• A single scene: the AI video scene prompt template.
• Animating an image you already have: the image-to-video template.
• Planning several shots before generating: the AI video storyboard template.
• Keeping a character or look consistent across clips: the consistent scenes template.
• Cutaways and landscapes: the b-roll and landscapes template.
• Ads: the 15-second product ad template and the UGC-style ad template.
• Short-form social video: the shorts, reels and TikTok script template.
• Online store products: the e-commerce product video template.
• A specific model: the prompt generator for Veo, Kling, Seedance, Wan, Hailuo or Runway, which writes in that tool's syntax.

Common mistakes

• Packing several actions or scenes into one short clip.
• Empty adjectives ("stunning", "premium look") instead of specific light, shot and style. Kling's guide lists this as a typical error.
• Writing negatives where they don't work: in Runway they can backfire and Kling 4.0 advises against them.
• Using one model's syntax in another, such as Hailuo 02 brackets in H3.
• Asking for long, exact on-screen text: Wan and Seedance list it among their typical errors.
• Changing the character description between clips: repeat it word for word and use reference images where you can.

Before you publish: don't imitate real people, their voice or likeness, or copyrighted characters, brands or music, and label the video as AI-generated where the platform or the law requires it. Model features and limits change often, so check what your tool supports today.