Writing scripts for AI video — what changes

In normal filmmaking the director reads the script and makes interpretive choices. AI reads literally. That changes the text you write.

Single scene prompt structure

1. Subject — what is in the frame: "30-year-old woman in red coat"
2. Action — what they do: "walks down a narrow street, glances over her shoulder"
3. Setting — where: "autumn Budapest, wet cobblestones, street lamps"
4. Camera — how it is shot: "medium shot, slow tracking, slightly handheld, eye level"
5. Lighting — light: "golden hour, warm rim light, soft shadows"
6. Style — overall aesthetic: "35mm film, Kodak Portra, slight grain"
7. Negative — what should NOT appear: "no text, no logos, no people in background"

Decomposition

One long prompt for a 30-second clip will not work. AI loses coherence past 5-8 seconds. Approach: write 5-second scenes, stitch in an editor.

• Each scene = its own prompt
• Continuity description between scenes: "same outfit", "same lighting"
• If a character appears across scenes — character reference image (available in Sora 2, Runway, Kling)

What you don't write in normal scripts but AI needs

• Exact timing in seconds (1.5-second turn, 2-second pause)
• Color of clothing and props specified word by word
• Camera type and lens — affects style
• Era and time period — otherwise AI defaults to the 2020s

Practical workflow

1. Start with a normal script — as if for a DP
2. Break into 3-8 second scenes
3. Convert each scene into a technical prompt (ChatGPT or Claude can structure it)
4. Generate 3-5 variants per scene, pick the best
5. Stitch in Premiere/Final Cut/CapCut