What to write, in what order, and which of it the model will actually act on.
When picture and audio come out of one forward pass, the sound is not a layer you add afterwards. What changes in the brief when you write it first.
For about three years, prompting a video model meant describing a picture. Sound was somebody else's problem — a library cue, a voice session, a foley pass, all bolted on downstream. Then models started generating audio and picture jointly, in one forward pass, and most people kept prompting them exactly the same way.
The result is a very specific failure. The image is fine. The audio is a room tone and a vague swell, because that is what the model produces when nothing in the prompt tells it what should be audible.
The fix is not to add a sound sentence at the end. It is to write the sound first.
When audio and video are decoded from the same latent, they are not two outputs that happen to be synchronised. They are one output. The model's estimate of what a scene sounds like is part of how it decides what the scene looks like — a line of dialogue implies a mouth, a mouth implies a framing, a framing implies a lens.
That causality runs in the direction most people do not expect. If you describe a wide establishing shot and then ask for dialogue, you have already told the model the mouth is 40 pixels tall, and it will render dialogue that fits that framing: mumbled, ambient, unresolved. If you write the line first and let the framing follow, you get the shot the line needs.
So the practical rule is: the loudest thing in the scene should be the first thing in the prompt.
The prompt shapes that hold up are boring and explicit. Three fields, in this order:
1. Audio event. What is heard, and by whom. A line of dialogue in quotes. A specific mechanical sound. A silence that is doing work. Name the language if there is speech — models with multilingual dialogue support usually have a short list they are genuinely stable in, often around a dozen, and a language outside that list degrades quietly rather than failing loudly.
2. Subject and action. Who is making the sound and what they are physically doing while making it. This is where you spend your specificity budget. "A woman in a wet raincoat pushes the door with her shoulder because her hands are full" beats any adjective pile.
3. Camera and frame. Last, not first. The camera is a consequence of the first two fields, and writing it first over-constrains everything downstream.
Most of the prompt-writing advice in circulation inverts this — it starts with "cinematic wide shot, 35mm, golden hour" and appends the content afterwards. That ordering was correct for silent image-sequence models. It is actively wrong for joint audio-video ones.
The most underused instruction in this whole category is the explicit absence of sound.
Models default to filling the audio track. Ambience, breath, a little music-shaped texture. If you want a beat of real silence — the horror cut, the held reaction, the moment before the line — you have to say so, and you have to say what replaces the sound. "The room tone drops out entirely for the length of the shot" produces a different and much better result than omitting any audio description at all, because the first is an instruction and the second is a vacuum.
Same for near-silence. "Only the refrigerator hum" is a strong prompt. Nothing is a weak one.
If you are evaluating a model, or evaluating your own prompt template, the cheapest useful test is a single spoken line on a single static shot.
It exercises almost everything that is hard: lip synchronisation, phoneme timing, prosody, whether the voice matches the apparent age and body of the subject, whether the room acoustics match the visible space, and whether the model holds a face stable while a mouth moves. A camera move or a scene change hides all of that behind motion. A locked-off shot with one line hides nothing.
Run that test at the shortest duration the model allows and the lowest resolution it offers. You are testing the audio-visual bind, not the pixels, and it costs a fraction of a full-length take.
None of the visual craft goes away. Framing, lens language, blocking, light — all still load-bearing. What changes is that they are now downstream of an audio decision, the same way they are on a real set where the script exists before the shot list.
That is the actual mental model. You are not writing an image description any more. You are writing a very short scene, and the scene has a soundtrack that the model is going to invent if you do not specify it.
If you want to see the field structure laid out rather than described, a page of briefs written sound-first is a faster way to internalise it than reading another list of adjectives — it makes the ordering visible, which is the whole point. The same structure applies whether you are working from a text brief or from a still, though the text-to-video path is where the audio-first ordering matters most, because there is no image to anchor the framing for you.
Write the line. Then the body. Then the lens. In that order, the audio stops being the thing you fix in post and starts being the thing the shot was built around.