LARUNDStart free
← All articles

GUIDE · 2026

How to write AI video prompts that actually work

Most weak AI videos come from weak prompts, not weak models. A prompt like “a cool video of a city” leaves the model to decide everything, so you get a generic result and no way to steer it.

This guide gives a repeatable prompt structure, a vocabulary for camera and motion, and examples of vague prompts rewritten into useful ones. It applies to most text-to-video and image-to-video models. Where we mention tools, Larund is the studio we build.

The five-part prompt structure

A reliable video prompt answers five questions in a fixed order. Models read the beginning most strongly, so put the most important information first.

  1. Subject: who or what is on screen, with two or three concrete traits.
  2. Action: what the subject does, in one clear verb phrase.
  3. Setting: where and when, including light and weather.
  4. Camera: the shot size, angle, lens feel and camera movement.
  5. Mood and style: the overall feel, such as calm documentary or high-energy commercial.

Describe motion, not just a picture

Video prompts differ from image prompts because something has to happen. A still description such as “a woman in a red coat on a street” gives the model nothing to animate, so it improvises subtle drift or a random action.

State what moves and how: “walks slowly toward the camera”, “steam rises from the cup”, “the wind moves her scarf”. Keep it to one main action per shot. Asking for several unrelated actions in a short clip usually produces a blur of half-finished movements.

Speed words help: slowly, quickly, gently, abruptly. So does stating what stays still, for example “the background stays static”, when you want only the subject to move.

Use camera and lens language

Camera terms are the strongest steering tool because they map onto how real footage looks. Shot size (wide, medium, close-up), angle (low-angle, high-angle, eye level), and movement (static, pan, tilt, dolly in, tracking, handheld) all change the result noticeably.

Lens feel matters too. A wide lens exaggerates depth and feels immersive; a longer lens compresses space and isolates the subject. In Larund, tokens such as &35mm and &low-angle pin those choices in the prompt, so the same shot reads the same way across scenes.

You wantTry wording
Establishing feelwide shot, static camera, slow push-in
Intimacyclose-up, shallow depth of field, eye level
Power or heroismlow-angle, slight upward tilt
Documentary realismhandheld, natural light, medium shot
Smooth product revealslow dolly in, soft studio light, clean background

Before and after: fixing vague prompts

The easiest way to learn is to compare. Each rewrite below keeps the same idea but adds a subject, an action and a camera.

Vague: “A cool video of a coffee shop.” Better: “Close-up of a barista pouring milk into a latte, steam rising, warm morning light through a window, slow dolly in, calm and cosy.”

Vague: “A car driving fast.” Better: “Low-angle tracking shot of a red sports car driving along a coastal road at sunset, wheels spinning, background blurred by speed, cinematic and bold.”

Vague: “A person talking about my product.” Better: “Medium shot of a woman in a kitchen holding a blue water bottle and smiling at the camera, soft daylight, handheld, natural and friendly.”

Text-to-video versus image-to-video prompts

With text-to-video the prompt must carry everything, so all five parts matter. With image-to-video the image already fixes the subject, setting and style, so the prompt should concentrate on action and camera. Repeating the whole look can conflict with the image and cause the model to redraw it.

Image-to-video is also the safer route for identity: a character or product that must look exactly right is better supplied as a picture than described in words. Larund builds video this way by default, starting from keyframes you approve before the motion is rendered.

Common mistakes

Most prompt problems fall into a short list. Fixing them often matters more than switching models.

  • Too many ideas in one shot; split them into separate shots.
  • Vague adjectives such as “beautiful” or “epic” with no concrete detail.
  • Contradictions, such as “static camera” together with “fast tracking shot”.
  • Unrequested text and logos; if you need exact text, add it afterwards with titles or motion graphics instead.
  • Long negative lists; say what you want rather than what you do not.
  • Changing everything between attempts; change one element at a time so you learn what mattered.

Iterate cheaply

Treat the first generations as drafts. Test wording on short, low-resolution clips before paying for a long premium render, and change one variable per attempt.

Different models also respond differently to the same wording. Because Larund has one balance for 25+ video models and shows the price before you generate, you can run one prompt on two models and keep the better result. For reference, a short video on Larund means a 6-second 480p clip on the cheapest model; premium models and 1080p cost more per second.

Is Larund the right pick?

A good fit when

  • You want to try the same prompt on several video models without opening several accounts.
  • You want lens and angle direction pinned with tokens such as &35mm and &low-angle.
  • You want to see the credit price before each generation while you iterate on wording.

Probably not the best fit when

  • You only use one video model and already know its prompt style well.
  • You need a prompt library tied to a single vendor’s proprietary controls.
  • You want a desktop editor to cut and composite clips frame by frame.

Frequently asked questions

How long should an AI video prompt be?

Usually 30 to 80 words. Long enough to cover subject, action, setting, camera and mood; short enough that the model does not drop details or follow contradictory instructions.

What is the best structure for a text-to-video prompt?

Subject, action, setting, camera and mood, in that order. Put the most important element first and keep one main action per shot.

How do I control the camera in an AI video prompt?

Name the shot size, angle and movement, for example “low-angle tracking shot” or “slow dolly in”. In Larund you can pin the framing with tokens such as &35mm and &low-angle.

Should I use negative prompts in AI video?

Often they are not needed. Describing what you do want is usually more reliable than listing what you do not, and some models ignore negative prompts entirely.

Why does my AI video ignore parts of my prompt?

The prompt may be too long, contain contradictions, or ask for several actions in a short clip. Reduce it to one subject and one action and add detail gradually.

Is there a free way to practise AI video prompting?

Larund’s Free plan offers 5 credits a month and 20 one-time welcome credits, about 4 short videos. Free videos have a small larund.com watermark.

Do prompts work the same across different AI video models?

The structure carries over, but each model has its own strengths and quirks. Test the same prompt on two models and keep wording that works on both.

Every top model. One studio.

Start free with Larund Studio: 25+ video and 30+ image models, keyframe films and motion graphics, with the price shown before every generation.

Try Larund free