How to write AI video prompts that work
A video model can only make what your prompt makes clear. This guide shows a simple structure for prompts, gives examples you can adapt, and explains why clips glitch and how to fix them. It applies to text-to-video and image-to-video in Frenz AI, and to most video models.
A structure that works
Write the prompt in this order. Each part answers a question the model would otherwise guess at.
- Subject — who or what, described by what is visible: "a woman in a yellow raincoat".
- Action — one clear movement: "walks toward the camera, then stops".
- Setting — where and when: "a wet city street at night".
- Light and mood — "neon reflections, soft rain".
- Camera — "slow dolly in, eye level, shallow depth of field".
Example prompts
- "A golden retriever runs along a beach at sunset, waves breaking behind it, low tracking shot, warm light."
- "Close-up of hands pouring latte art into a white cup, steam rising, soft window light, static camera."
- "A paper boat drifts down a rain gutter, slow push-in, overcast daylight, shallow depth of field."
Why clips glitch — and the fix
- Too many actions in one clip → keep one main action per clip.
- Hands, text and small details warp → keep them out of focus, or out of frame.
- Faces drift between frames → use image-to-video with a clear photo and reference images.
- Busy crowds smear → fewer subjects and a simpler background.
- Fast camera moves tear the image → slow, single camera moves.
Work in small steps
Test a prompt at the shortest length first. When it looks right, make the longer, higher-quality version. Changing one thing at a time tells you which change helped.
Questions people ask
How long should a prompt be?
Usually one to three sentences. Longer prompts are allowed, but clarity matters more than length.
Should I describe what I don't want?
State what you do want instead. "A still lake" works better than "no waves".
Related
Ready to start?
Last updated