How to Write Prompts for AI Video Models

How to write prompts for AI video models: a working structure of subject, motion, camera, light and duration — plus the five mistakes that waste the most credits. Better prompts for AI video models cost nothing and double your hit rate.

Updated September 3, 2026

How to Write Prompts for AI Video Models: what this guide covers

Use five slots, in this order
Name the camera, not the vibe
One motion per clip
Describe light like a gaffer would
The five mistakes that waste the most credits
Reuse what works

Use five slots, in this order

Subject, motion, camera, light, then constraints. 'A cyclist (subject) pedalling hard along a coastal road (motion), tracking shot from a low angle (camera), golden hour backlight (light), 5 seconds, 16:9 (constraints).' Models read this order well because each slot narrows the previous one. Putting adjectives before the subject — 'a breathtaking, cinematic, gorgeous cyclist' — spends tokens on words that change nothing in the output.

Name the camera, not the vibe

'Cinematic' is not a camera. '35mm lens, slow dolly in, shallow depth of field' is. Camera language is the single highest-leverage thing you can add to a video prompt, because it is the part models are genuinely trained on. If you only improve one thing about your prompts, make it this.

One motion per clip

Text-to-video models are reliable for one camera move and one subject action per clip. Ask for a dolly in plus a pan plus a subject turning plus a background change, and you will get four motions smeared into something incoherent. If a shot needs more, generate two clips and cut.

Describe light like a gaffer would

Light direction and quality beat colour adjectives. 'Low-key practical light from a single window, hard shadows' produces a recognisably different image from 'moody lighting'. Adding the light source — window, neon sign, overcast sky — gives the model a physical reason for the look, and models reward physical reasons.

The five mistakes that waste the most credits

First: stacking style adjectives that contradict each other. Second: asking for text to appear on screen — most models render garbled lettering, so add titles in your editor. Third: ignoring duration and getting a clip that is too short to cut with. Fourth: regenerating the same prompt on the same model hoping for a different result, when switching model would be cheaper. Fifth: using a flagship for a draft. Drafts belong on fast models.

Reuse what works

When a prompt produces something good, save it as a recipe with the model, duration and aspect ratio attached. Prompt quality compounds only if you keep the prompts that worked — otherwise you rediscover the same phrasing every month.

Models mentioned in this guide

Veo 3.1

Google · 85 credits · 8s max

Wan 3.0 Prime

Alibaba · 60 credits · 15s max

Hailuo 2.3

MiniMax · 52 credits · 10s max

Seedance 2.0

ByteDance · 72 credits · 12s max

How to Write Prompts for AI Video Models FAQ

How long should a video prompt be?

One to three sentences is the sweet spot. Longer prompts do not add control; they add contradictions. If you need more control, move to a model with reference inputs rather than writing a longer paragraph.

Can I make text appear in the video?

Not reliably. Most video models render lettering as plausible-looking gibberish. Generate a clean plate and add titles in your editor.

Why does the same prompt give different results?

Generation is stochastic — models sample from a distribution, so the same prompt legitimately produces different clips. Run it two or three times and pick, rather than assuming the first output is what the prompt means.

More guides

How to Choose an AI Video Model Text-to-Video vs Image-to-Video: Which Should You Use?AI Video Resolution and Duration: What Each Model Actually Supports AI Video Models With Native Audio: The Complete List How to Compare AI Video Models Properly