What you’ll learn
- A five-part prompt structure (subject, action, setting, style, camera) you can reuse every time
- How to rewrite a vague prompt into a precise one — with before/after examples
- Why negative cues (“no text, no watermark”) clean up your results
- A one-change-at-a-time iteration loop that improves shots without guesswork
How models read your prompt
A text-to-video model doesn’t “understand” a story the way a person does — it pattern-matches your words to visuals it has learned. The more concretely you describe what should be on screen and how the camera behaves, the less it has to guess. Specific nouns, one clear action, and a named camera move beat a pile of mood adjectives.
A prompt structure that works
Describe the shot the way a director would — subject, action, setting, style and camera. Keep it to one or two clear sentences.
- 1Subject: who or what is on screen (“a barista”, “a sleek phone”).
- 2Action: what happens (“pours latte art”, “rotates slowly on a pedestal”).
- 3Setting: where (“sunlit specialty café”, “minimal studio, soft shadows”).
- 4Style: the look (“cinematic, shallow depth of field, warm grade”).
- 5Camera: the movement (“slow push-in”, “orbit”, “static wide”).
Before and after
Vague: “a coffee video”. Better: “Close-up of a barista pouring latte art into a white cup on a wooden counter, sunlit specialty café, cinematic, shallow depth of field, slow push-in, no text.” The second prompt controls subject, setting, lighting, lens and motion — so the model has far less to invent, and you get a usable shot more often.
Say what you don’t want
Negative cues clean up results. Add “no text, no watermark, no logos, no extra fingers” to avoid the classic AI artefacts. For brand-safe output, add “generic packaging, no brand logos”. A short list of exclusions often does more for quality than another adjective.
Iterate, don’t over-specify
Start with a focused prompt, generate, then change one thing at a time — the lighting, then the camera move, then the mood. Stacking ten adjectives at once makes it impossible to tell what actually helped. Treat each generation as an experiment with a single variable.
Build a reusable “house style”
Once you find a look you like, save the styling half of the prompt as a suffix (e.g. “cinematic, 4K, soft natural light, shallow depth of field”) and reuse it across clips. You change the subject and action per shot, but the house style keeps a series visually consistent — which is what makes a channel look intentional.
Quick tips
- Lead with the most important visual — models weight the start of the prompt more heavily.
- Name the camera move (“slow pan”, “orbit”, “static”) to control energy and pacing.
- Reuse a “house style” suffix for a consistent look across every clip.
- One action per shot — split “walks in and sits and talks” into separate scenes.
- Keep a swipe file of prompts that produced great shots; reuse and remix them.
Frequently asked questions
How long should a prompt be?
One or two clear sentences usually beats a paragraph. Be specific, not long.
Can I use the same prompt across models?
Yes — on Vivideo you can run one prompt through different models (Sora, Veo, Kling and more) and compare.
Why does my video ignore part of the prompt?
Models prioritise the start and can drop later details. Move the key element earlier, or split it into another scene.
Do prompts work for image-to-video too?
Yes — with an input image the prompt mainly directs motion and camera, not the subject.
How do I keep a consistent character across shots?
Reuse the same detailed subject description, or use an avatar/reference image so the look stays stable.