How-to · 4 min read
How to write better AI video prompts (a working guide)
The single biggest difference between great AI video and mediocre AI video is the prompt. Most beginners write what they want to see ('a dragon flying'). Pros write the entire shot: camera, subject, lighting, mood, motion. This guide is the prompt formula that consistently outperforms the obvious version.
What you need
- Any AI video model
Steps
Start with the camera, not the subject
Bad: 'A dragon flying.' Better: 'Slow tracking drone shot, golden hour. A dragon made of fire flies over a frozen fjord.' The camera language tells the model what kind of shot you want.
Specify lens and depth of field
Add 'shot on 50mm, shallow depth of field' or '24mm wide angle' or 'macro close-up'. These map directly to how the model renders the frame.
Describe lighting like a DP
'Golden hour', 'overcast morning', 'neon-lit night', 'single key light from the left, deep shadows'. Lighting is the #1 lever for mood.
Pick a motion verb
'Dolly in slowly', 'rapid pan', 'static locked-off shot'. The motion verb tells the model how the camera should behave.
Add a film/look reference where allowed
'Shot in the style of Blade Runner 2049 cinematography' or '35mm anamorphic'. Models trained on broad data recognize many references.
Keep it one strong subject per shot
Models break when you stack too many concurrent actions. Pick one subject, one motion, one mood per generation. If you need more, generate multiple clips and stitch them.
Notes
- Sora 2 listens to camera grammar. Veo 3.1 listens to lighting language. Kling 3.0 listens to motion verbs. Tune the prompt to the model.
Start with VIBE
19 AI video models including Sora 2 and Veo 3.1. Free starter generations.
FAQ
- One to three sentences for most models. Sora 2 Pro and WAN 2.6 can handle longer detailed prompts.