When to use it
You need a short clip from text (a title card, b-roll, a concept scene) in Sora/Veo/Kling/Runway, but the generations jitter and drift off. The AI's role is a director/cinematographer. Output: a prompt that specifies the scene, camera and subject movement, lighting and duration — the model gets a storyboard instead of a wish.
The prompt (copy and paste)
You are a director and cinematographer. Build a prompt for the video generator <PASTE: Sora / Veo / Kling / Runway>.
CLIP IDEA: "<PASTE: what happens on screen>".
DURATION: <e.g. 5 seconds>. FORMAT: <16:9 / 9:16>.
MOOD/GENRE: <e.g. cinematic, warm sunset>.
Describe ONE continuous scene, point by point:
1. Subject and action (what it does, slow/fast).
2. Environment and light (where, time of day, light source).
3. Camera movement (static / slow push-in / orbit / tracking the subject) — ONE main movement.
4. Style (cinematic 35mm / anime / 3D), pacing and atmosphere.
5. What changes on screen over the clip (start → end) — a small arc, no change of location.
One coherent description. No hard cuts and no scene changes inside the clip. End with "Avoid:" (flicker, distorted faces/hands, on-screen text).
The technique: video models really need ONE camera movement and a short "start→end" arc within a single scene — that's what removes the jitter and the abrupt, illogical cuts.
Filled-in example
Idea: "a drop of milk falls into coffee". Duration: 5 sec, 9:16, cinematic. Platform: Veo.
What the AI should come back with (roughly): "Extreme close-up of a single drop of milk falling into a cup of black coffee, slow motion, soft warm window light from the left, camera slowly pushes in on the surface, cinematic 35mm, shallow depth of field; over the clip the drop hits and a gentle crown-splash ripples outward. 9:16, calm premium mood. Avoid: flicker, distortion, on-screen text, scene changes."
Variations
- Bring a still to life (image-to-video). "Describe the minimal motion for this frame: what comes slightly alive, where the camera drifts" — for animating a static image.
- A 3-clip storyboard. "Split the story into 3 prompts of 5 seconds each so I can cut them together" — long videos are assembled from short scenes.
- Vertical for Shorts/Reels. Strictly 9:16, subject centered, action shot large — built for the mobile feed.
Pro tips
- One camera movement per clip. "Push-in + orbit + pan" all at once is mush; shoot complex moves as several clips and join them in the edit.
- Keep ONE scene inside a generation: models handle location changes badly (illogical morphs). If you need a transition, that's two clips.
- Faces and hands are the weak spot of video models: for close-ups of people plan for extra takes and simple poses, and explicitly ask for no distortion in the "avoid" block.