AI Video Prompts: Describe the Motion, Not Just the Scene

Here's a prompt most people write on their first try:

A woman in a red coat on a rainy city street at night.

It describes a photograph. A video model will give you something, but you've left the motion up to chance: she might stand still, glide sideways or melt into the pavement. Now the same idea with the motion written in:

Medium tracking shot, a woman in a red coat walking toward the camera along a rainy city street at night, her hand lifting to pull the collar up against the wind. Rain streaks through the glow of shop signs, puddles ripple under each step. The camera slowly backs away at her walking pace.

Same scene. The second one tells the model what the subject does, what the camera does and how fast. That's the whole skill behind good ai video prompts, and this post breaks it down.

Why a scene description isn't enough

An image prompt only has to describe one frame. A video prompt describes change over time. If you don't say what changes, the model fills the gap with its own guess, and the guess is often a slow drift or a stiff, barely moving clip.

The model makers' own guides point the same way. Google's Ultimate prompting guide for Veo 3.1 uses a five-part formula: cinematography, subject, action, context, style and ambiance. Action has its own slot. Their example is a tired worker "rubbing his temples in exhaustion", a specific visible movement rather than "looking tired".

Kling's prompt guide lists subject, action, scene, camera and lighting or mood, and tells you to describe what a viewer would actually see: "smoke drifting upward", "flames bending in wind", "runner leaning forward".

Two different companies, one habit: name the motion in plain, visible terms.

Three layers of motion

When you write a video prompt, think of three things that can move. Most weak prompts cover none of them.

1. Subject motion

What does the person, animal or object physically do? Use verbs you could film.

• Weak: a chef cooking
• Better: a chef tosses diced onions in a steel pan, flames flaring up, then tips the pan to slide them onto a plate

Avoid abstract emotion words on their own. The fal.ai LTX-2.5 guide suggests expressing character through physical detail, with examples like "a jaw held tight, a hand that won't stay still", instead of just saying someone is nervous.

2. Environment motion

Rain, steam, dust, fabric, water, crowds in the background. These small movements make a clip feel alive, and they're cheap to add: steam curls off the coffee, curtains shift in a light breeze. Kling's examples (smoke, flames) are this exact layer.

3. Camera motion

Say what the camera does, using terms filmmakers use. Google's guide lists dolly shots, tracking shots, crane shots, aerial views, slow pans and POV shots. Kling's guide adds push-ins, pans and tilts for reveals, and tracking for action continuity.

Their crane example is a good model of how much detail helps: "Crane shot starting low on a lone hiker and ascending high above, revealing they are standing on the edge of a colossal, mist-filled canyon at sunrise." The camera move and the reveal are both in the sentence.

A simple formula you can reuse

Both Google and Kling use a version of this. Mine is a blend:

1. Shot and camera: Slow dolly-in, medium close-up
2. Subject: a barista in a green apron
3. Action: pours milk into a latte, drawing a leaf pattern
4. Setting: in a sunny corner café, morning
5. Style and light: warm natural light, shallow depth of field

Put together: Slow dolly-in, medium close-up, a barista in a green apron pours milk into a latte, drawing a leaf pattern, in a sunny corner café in the morning, warm natural light, shallow depth of field.

You don't need every slot every time. But if you only fill one, make it action.

Pace matters as much as direction

"Moves" isn't enough. Does she stroll or sprint? Does the camera creep or whip? Kling's guide suggests pace words such as "slow push-in" and "steady tracking shot". Add slowly, at walking pace, in one smooth motion, abruptly. Vague camera direction is a common reason for inconsistent results, and a speed word is the cheapest fix.

Mistakes people actually make

• Stacking camera moves. Orbit, zoom in, crane up and pan left in one short clip invites warping. Pick one move per shot. Fast, complex camera work tends to be where models fall apart.
• Describing the scene twice. Long lists of adjectives about the setting use up the prompt while the action gets one vague verb. The LTX-2.5 guide even ranks the extras, calling dressing the first thing to trim when a prompt sprawls.
• Cramming a story into one clip. Five actions in five seconds usually means none of them looks right. One subject, one main action.
• Rewriting the image in image-to-video. If you start from a still, the picture already shows the subject, outfit and setting. Spend your words on what happens next. The fal.ai guide shows the other side of this: when you want a locked shot, say so, for example "Hold the decanter, the bar top, the backlight and the framing of the still exactly as they are."
• Expecting readable text. Signage and on-screen text remain unreliable in several models, including LTX according to fal.ai. Add text in an editor afterward.
• Writing the camera move in a settings dropdown only. The fal.ai guide recommends keeping camera direction in the prose, because a written move can be tied to a specific moment in the action and a dropdown value can't.

Sequencing: when you want more than one beat

Some models can handle timed beats in one prompt. Google's Veo 3.1 guide shows timestamp prompting, with segments like [00:00-00:02] for a shot of an explorer pushing aside a vine, then [00:02-00:04] for a reverse shot of her face. Each segment gets its own camera and action.

Two cautions. First, this is a documented technique for Veo, not something every model supports, so check how your model behaves. Second, fal.ai's LTX guide says to re-establish framing at each cut, because "scale, angle, lens and light all reset at a cut and the model has nothing to infer them from." It also suggests 2 to 4 shots per generation as the sweet spot for that model. If a multi-shot prompt falls apart, generate the shots separately and edit them together.

Sound is part of the prompt on some models

Veo 3.1 generates audio with the video, and Google's guide suggests writing dialogue in quotation marks and sounds as lines like SFX: thunder cracks in the distance. The LTX-2.5 guide treats sound as the highest-priority element for that model and says each sound should have a source in the frame, such as a marshal at a post, rather than a mood word. If your model produces audio, give it something to sound like. If it doesn't, leave these lines out.

Testing it in practice

Results still vary from run to run, even with a perfect prompt. Treat the first generation as a draft.

1. Write the shot with one subject action and one camera move.
2. Run it and watch for the specific failure (stiff subject, wobbling camera, wrong speed).
3. Change one thing: the verb, the camera term or the pace word. Not all three.
4. Keep the version that works in a saved list of prompts.

In Mediafy Studio you can pick a video model such as Kling, Veo 3.1 or MiniMax, type your prompt, and check the credit cost on the Generate button before you run it. If your prompt is short, the Enhance button rewrites it into a fuller one, which is a decent starting draft to edit by hand. Check that the motion verbs survived. Mediafy Chat can also help you break a idea into shots. Different models read the same prompt differently, so running one prompt on two models is a quick way to see which handles your kind of motion better.

If you want to start from a still, make the image first (our first image tutorial covers the basics) and then write a prompt that only describes the movement. For a structured walk-through of the whole workflow, there's also the AI Creator Lab course.

A quick checklist

• One subject, one main action, written with a verb you could film
• One camera move, with a speed word
• One or two small environment movements
• No text you need to be readable
• Sound described only if the model makes sound
• Change one variable per retry

Ready to try it? You can sign in with Google and run your first video prompt.

Sources

• Google Cloud Blog, "Ultimate prompting guide for Veo 3.1": https://cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-veo-3-1
• Kling AI, "Kling AI Prompt Guide: The Secret to Cinematic Video Prompts": https://kling.ai/blog/kling-ai-prompt-guide
• fal.ai, "How To Use LTX-2.5: Prompts & Workflows [2026]": https://fal.ai/learn/tools/how-to-use-ltx-2-5