I always write prompts in English, even when the video is for an Arabic-speaking audience. Video generation tools understand English better, and the result is more accurate. This is not translation — it means the technical language of the prompt stays English.
Step one: the reference image
Before writing a single word, I upload an image that sets the frame: camera angle, lighting, subject, environment. The reference image is half the result. The text prompt describes motion and feeling, but it does not invent the look from nothing.
That is why I keep a standing folder of references — street shots, night lighting, close faces. With references ready, the text prompt gets shorter and sharper.
Step two: the three-part structure
In the text I describe three things, in this order:
- Motion: what happens — who moves, where, and how fast.
- Visual style: lens, lighting, colour grade, shot size.
- Mood: the feeling the first two seconds should carry.
I avoid long sentences and vague adjectives. “A man walks slowly down a wet street at sunset” beats “a cinematic dramatic scene of a character moving through an urban environment.”
Step three: structured iteration
I test the same prompt three or four times, changing exactly one thing each time: a different angle, different lighting, or a different motion speed. Changing one variable tells me why a result improved; changing everything at once wastes the lesson.
The goal is not the first acceptable result — it is learning which description gives the tool the best chance. That knowledge compounds and becomes the fastest asset you own.
Mistakes that ruin the output
- Describing a whole story in one prompt; the tool renders a shot, not a film.
- Stacking contradictory adjectives like “calm, chaotic and cinematic”.
- Relying on text alone with no reference image.
- Asking for written words inside the frame; they usually come out distorted.
- Judging a tool from a single attempt.
After generation
Sound and effects come later; they are not part of the video prompt. AI generates the moving image, while pacing, music and shot order stay human decisions — and that is what separates a technical clip from one people watch to the end.
If you want a full worked example with the actual prompts, I broke down every shot of the RESILIENT video in the tutorial.