Long read

Prompting AI Video: What Carries Over

A video prompt is an image prompt plus time. The walls still do the styling. The new work is motion, camera, and sound.

Free Veo Sora Kling 3 min read

What transfers from image prompting

Almost everything. The video models learned from the same visual language the image models did, so the walls keep working. Golden hour lights a clip the way it lights a still. Ukiyo-e styles the frames. Monsoon soaks the scene. Write the visual half of a video prompt the way you write an image prompt: subject, style, lighting, weather, mood.

The habits transfer too. Specific beats general, one style at a time, name what you want in the frame. If your image prompts work, the front half of your video prompts already works.

What video adds

One action per shot. This is the biggest single rule. A fox trots across fresh snow renders clean. A fox trots, catches a mouse, looks up, and runs to its den smears into morphing. Ask for one motion and let the clip breathe.

Camera moves are words now. The composition wall's camera terms become instructions the model executes over time: slow dolly in, orbit around the subject, static shot, handheld. Pick one move per shot, the same way you pick one action.

Sound is part of the prompt. The current audio-capable models take dialogue and effects inline: she whispers "we found it", wind through the pines, distant thunder. Write sound the way you write light.

Time words matter. A single slow beat, time-lapse, loop tell the model how the seconds should spend themselves.

The models, in one paragraph

Veo lives inside Gemini and Flow and takes dialogue and sound effects in the prompt. Sora lives in the ChatGPT world and favors plain sentences. Kling and Runway sit closer to the classic generator workflow, with strong image-to-video modes that animate a still you already made. That last path is the easiest bridge from this site: build the image with the walls, then hand it to a video model with a one-line motion instruction.

Common failures and their fixes

Morphing. Too many actions or too many subjects. Cut the prompt to one subject, one action, one camera move.

Style drift. The clip starts as watercolor and ends as photography. Lock it by naming the style twice: once up front, once as a closing phrase like consistent watercolor style throughout.

Teleporting. The subject jumps position between frames. Give it a path instead of a destination: walks left to right along the fence line.

Dead air. A clip with no motion at all reads as a broken image. If the subject holds still, move the world: fog drifts past, candle flame flickers.

Worked prompts

scenea lighthouse on a rocky headland at dusk, waves breaking below, slow dolly in, golden hour, cinematic, one continuous shot
charactera Kitsune trots across a moonlit shrine courtyard, lanterns swaying, static wide shot, ukiyo-e style throughout
ambient looprain on a Tokyo alley at night, neon reflections in puddles, steam rising, locked-off shot, seamless loop
with soundan old keeper lights the lamp and says "one more winter", gulls crying, wind against glass, handheld, warm tungsten light

Published here first.