One scene, three AI video prompts: LTX, MiniMax H3 and Kling

Three AI video models now generate picture and sound together, cut between shots inside one clip, and put words in a character's mouth. Each publishes a guide to writing for it, and the three guides disagree about the form a scene should take. Lightricks says it plainly: a prompt written for another model loses its tag syntax and shot-list formatting on the way into LTX, and tends to underperform. What follows is one scene written three ways, from those guides.

The scene: a woman waits at a rainy bus stop at night; she checks her phone and says "He's late"; a bus pulls in.

LTX: write it as prose

Lightricks wants a single flowing paragraph in the present tense — roughly four to eight sentences for a single take — with dialogue in quotation marks and emotion shown through physical cues rather than named. For LTX-2.5's multi-shot scenes the cut goes into the prose too: name the transition, re-establish the new framing, say whether the sound carries across. It recommends two to four shots per generation.

A wide shot frames a bus stop at night, rain streaking through a streetlight's cone. A woman in a grey coat stands under the shelter, glancing at her phone; its light catches her face. She exhales and says quietly, "He's late." A hard cut moves to a close-up of the phone screen as rain ticks on the shelter roof; the ambience continues across the cut. Headlights sweep across the glass as a bus pulls in.

One more line in the guide changes how a writer paces a scene: LTX-2.5's duration predictor times the clip to the action as written, and will not add a pause you did not prompt for. If the beat matters, write "a beat of silence" into the scene.

MiniMax H3: a tagged script

MiniMax's prompt guide for H3 is the most like a screenplay, and the strictest. The prompt has three named fields: integrated_multimodal_description for everything seen and heard in order, overall_soundscape for ambience and physical sound, and non_diegetic_music for score the characters cannot hear. Shots are numbered; every shot after the first opens with its cut time; a speaker gets a stable ID and their line goes inside a dialogue tag with the language named.

integrated_multimodal_description: [Shot 1] Live-action, cinematic, a wide shot frames a woman in a grey coat under a bus shelter at night as rain falls past a streetlight. She glances at her phone, and the tired woman with a low, quiet voice (S1) says: <d>[English] He's late.</d> [Shot 2] At 00:04.000, the camera cuts to a close-up of the phone screen as headlights sweep across the shelter glass.

overall_soundscape: Steady rain drums on the shelter roof and runs off the gutter. A bus engine approaches and its brakes hiss as it stops. non_diegetic_music: N/A

The guide also asks for camera moves written as sentences with an amplitude and a speed — "the camera pushes in with small amplitude at slow speed" — rather than labels stacked at the end, and keeps on-screen text in double quotation marks exactly as it should appear.

Kling: a timeline and keyframes

Kling 4.0 takes up to ten keyframes and prompts of up to 8,000 tokens, and its keyframes guide lays a prompt out as a timeline: 0–5s:, 5–10s: and so on, each segment saying what happens. The keyframes fix the important moments; the prompt, in Kling's words, explains how the video moves between them. For this scene that is two keyframes — the woman at the stop, the bus arriving — and a timeline that carries her from one to the other.

0–4s: wide shot, a woman in a grey coat waits under a bus shelter at night in heavy rain; she checks her phone and says quietly, "He's late." 4–8s: headlights sweep across the shelter glass as a bus pulls in and stops beside her.

Which to write first

Write the scene once as prose — it is the form closest to how a writer thinks, and it is what LTX wants anyway. Converting prose into H3's fields is mechanical: split it at the cuts, give each speaker an ID, move the rain into the soundscape. Converting it into Kling's timeline means deciding which moments are keyframes, which is a directing decision rather than a formatting one.