higgsfield-video-prompt (video-gen prompt architecture)

The layer that turns an approved still keyframe + a one-line shot intent into a tight, engine-tuned motion prompt for Higgsfield. Engine-agnostic in shape, Seedance 2.0-tuned by default. Use it every single time we generate a clip, so prompts stop being rough guesses and follow one repeatable structure.

Step 0, Routing gate (run this FIRST, before writing any prompt)

Most ad clips stay right here: a short (≤~20s) live-action ad built from per-shot i2v clips stitched in ffmpeg, with VO + music added in post (that was F). But if the job trips ANY trigger below, STOP and load the seedance-advanced skill first, it carries the technique this everyday skill deliberately omits. This gate is deterministic: because this skill runs on every clip, the check can't be forgotten.

If the job needs…	Trigger test	Where in `seedance-advanced`
Long-form planning	total length ≥ ~20s OR more than 3 shots (45s / 1-min always)	§3-4 (shot-list, continuity, multi-shot / continuation)
On-screen speech (lip-sync)	a mouth moves + says a real line in-render	§2
In-engine sound	footsteps / ambience / SFX / beat-sync baked in (not post)	§2
Cuts inside ONE generation	real editorial cuts within a single clip	§3
Clip continues clip	one clip opens from another's footage / final frame (not ffmpeg stitch)	§4
2D / anime / cel look	storyboard-board style	§6
Multi-reference roles	image=identity, video=camera, audio=tempo	§7
Non-English on-image dialogue	spoken / on-frame non-English	§2, §8

Long-form note: even when we still stitch per-shot clips in ffmpeg (our usual method), a ≥20s / >3-shot ad needs the shot-list + continuity-lock discipline from seedance-advanced §3-4 at PLANNING time. The in-engine multi-shot generation is optional on top; the planning is not. None tripped → continue below.