CHAPTER V · BUILDER

Compose by the five-part structure.

A live form that assembles the prompt and request body for any of the four supported models. Pick a slot, watch it land in the JSON, copy it.

Field Guide / 2026.1 · Updated 2026·05·05

00 Prompt builder · all models

Build for any model.

Pick a target video model. The form adapts to that model's request schema and prompting conventions, then assembles the request body live so you can copy it straight into Replicate, fal, or your own API client.

00 · Target model

Seedance 2.0text-to-video, joint audio Wan 2.7 i2vimage-to-video, motion brief Veo 3.1 fasttext-to-video, scene + audio Kling 3.0 i2vimage-to-video, camera moves

Five-slot anatomy: subject · action · camera · style anchor · constraints. Time-code your shot blocks. Audio is generated jointly — describe the sound.

01 · Subject (noun + 2–3 traits)

02 · Action (one verb)

03 · Camera

04 · Style (one dominant anchor)

05 · Constraints (negatives as positives)

06 · Reference image URL (i2v models)

Wan and Kling are image-to-video. Drop a public URL or upload to Replicate/fal first. The image carries 90% of the look, so the prompt focuses on motion, not style.

06 · Negative prompt (Veo only)

resolution

aspect

duration (s)

seed

generate_audio joint audio + dialogue

gateway

output format

…

What changes per model.

Seedance 2.0. Five-slot anatomy. Time-coded shot blocks ([00:00-00:04]). Joint audio — describe ambient + dialogue + score policy. Aspect and duration in the JSON.
Wan 2.7 i2v. The image is the style. Prompt is a motion brief: what the subject does, where the camera moves, what enters frame. Skip the style anchor.
Veo 3.1 fast. Scene-first prose works. Use a negative prompt for the things you want suppressed. Audio is generated jointly; tag it like Audio: rain ticking, no music.
Kling 3.0 i2v. Strong with explicit camera moves (slow push-in, orbit left). Pro tier handles longer durations. Subject motion verbs; the image carries the look.