AI video prompting is more demanding than image prompting. You're not describing a single frame — you're describing motion, time, and change. Every element that can go wrong in an image becomes five times more likely to go wrong across 8 seconds of video. Structure is the only thing that keeps it consistent.
This article covers the structure that works: the bracket format, the motion-first principle, identity lock, camera specification, and duration.
An earlier version of this page said the same structure applies to every video model. Checked against the vendors' own documentation, that is not true. They publish three different structures, and one of them explicitly rules out brackets.
Seedance 2.5 — ByteDance publishes an eight-part formula, and it is not this one. Through Dreamina, their own platform, the stated order is Format → Subject → Action → Environment → Camera → Look → Timing → Audio & Constraints. Format comes first (duration, aspect ratio, shot structure) and action comes third, not first. Their template also binds references with @Image1/@Image2, puts shot size before style language, allows one camera move per shot, budgets roughly 6–8 seconds per beat, and places exclusions last — after a positive Preserve [identity, wardrobe, logo, object geometry] clause. See the full template.
Runway Gen-4/4.5 — full sentences in natural language, ordered subject → action → setting → camera → motion over time → style. Runway's guidance is to write prose, not fields. A bracket prompt is working against it.
Veo 3.1 — Google's published formula is [Cinematography] + [Subject] + [Action] + [Context] + [Style & Ambiance], which opens with the camera, not the motion — the reverse of the motion-first principle below. Veo also has its own audio syntax: dialogue in quotation marks, effects prefixed SFX:, background prefixed Ambient noise:.
Kling 3.0 — no published bracket requirement, and unlike the others it has a real negative-prompt field.
What does carry across all four is the thinking, not the syntax: name the motion explicitly, anchor identity so it cannot drift, specify the camera once, and state duration. Every vendor formula contains those four things somewhere — they disagree only on order and packaging.
So what is the bracket format below? It is a working structure that was tested on Seedance 2.0 in June 2026 and produces good output. It is not, and this page should not have implied it was, the structure ByteDance publishes. Use it as a checklist that guarantees you have not forgotten motion, identity, camera or duration — then order the finished prompt the way your model’s vendor asks. Per-vendor templates with sources: what the vendors actually say.
Vendor documentation: Luma and Melies Seedance guides · help.runwayml.com · Google Cloud Veo 3.1 prompting guide · Kling API docs. Checked 25 Aug 2026Why Video Prompts Need More Structure Than Image Prompts
In a single image, a vague subject description produces a generic image. In a video, a vague subject description produces a different-looking person in every frame. The model isn't just guessing the subject once — it's regenerating its interpretation of the subject across every frame, and without explicit anchoring it will drift.
Motion introduces another layer. The model needs to understand what's moving, how it's moving, at what speed, and from what camera angle — all simultaneously. Without clear structure, motion instructions bleed into subject descriptions, camera movements get conflated with subject movements, and the output becomes incoherent.
The Bracket Format
The bracket format organises video prompt elements into labelled sections. The model parses each section separately, which prevents element bleeding and produces significantly more consistent results.
You don't need all six sections for every prompt. Short lifestyle clips might only need [MOTION], [SUBJECT], [ENV], and [DURATION]. Complex cinematic sequences need all of them. The key is: whatever you include, put it in its own labelled bracket.
The Motion-First Principle
Motion should be the first thing in your video prompt — before the subject description, before the environment, before anything else. AI video models weight earlier tokens heavily during the generation process. If you put the environment first, the model prioritises environment. If you put motion first, the model prioritises motion — which is what you want.
Identity Lock — The Most Important Line in Any Video Prompt
Temporal consistency — keeping the subject looking the same across every frame — is the most common failure mode in AI video. Without explicit instructions, the model will subtly change the subject's face, hair, and clothing between frames. Over 8 seconds, this drift is obvious and unusable.
Identity lock is a prompt instruction — not a model setting. You add it explicitly to the [SUBJECT] section:
The phrase "no morphing" is particularly effective — it directly names the failure mode you're preventing. For models that support reference image upload (Higgsfield Studio, ControlNet-based pipelines), use a reference photo in addition to the text lock. Text alone may not be enough for high-movement sequences.
Camera Specification
The camera bracket controls how the scene is filmed — not what's in it. Separating camera instructions from subject instructions is essential. When they're mixed, the model often applies movement instructions to the subject rather than the camera.
Key camera movements for AI video:
Tracking dolly — follows the subject laterally. Produces the most natural-looking motion for walking sequences. Dolly push-in — moves toward the subject. Creates intensity and intimacy. Handheld — introduces authentic shake. Good for UGC and documentary aesthetics. Aerial drone — overhead or elevated perspective. Static tripod — zero camera movement. Forces the subject's motion to carry the scene.
Adding a lens reference (35mm anamorphic, 85mm portrait, 24mm wide) changes the optical quality of the output significantly — the model understands the depth, compression, and bokeh characteristics associated with each lens.
Duration and Frame Rate
Most AI video models support clips up to 10–15 seconds. Always specify duration in the prompt. Models will default to their shortest clip length if you don't specify, and you'll often get a truncated version of the motion you described.
"No scene cuts" is important. Without it, some models will insert a hard cut in the middle of the clip — which destroys the continuity you were trying to build. "Continuous motion throughout" reinforces that the action should flow from start to finish without interruption.
A Complete Working Example
More prompts. Every week.
Production-ready prompts, model guides, and AI workflow breakdowns — free forever.
SUBSCRIBE FREE ↗