Veo 3.1 Prompts for Campaign Film

Every video model before this one handed you a silent clip. You wrote for the eye, and sound was a separate job that happened later, with different people, in a different room. Veo 3.1 generates audio with the picture — ambience, effects, spoken dialogue — which means the sound is now part of the take. If you do not direct it, you will get whatever the model thinks a scene like yours sounds like.

That is a genuine shift in the craft, and most prompts written for silent models handle it badly. This guide covers how to direct sound, how to write dialogue that arrives usable, and the specific mistakes that turn a beautiful clip into one the editor cannot cut.

Sound Is Now a Directorial Decision

The first question to settle, before writing anything: does this clip carry its own sound, or does it arrive silent?

Both are legitimate and the choice is structural. A campaign film cut to a licensed track wants silent clips — generated ambience underneath a scored edit is the single most common reason a good-looking clip gets thrown away. A social piece designed to work with the sound on, or any clip with dialogue, wants native audio.

Decide this first, because it changes the whole prompt. It is not a setting you flip afterwards.

The Three Layers of Sound

Treat audio the way a sound designer would, as distinct layers, and name them separately. Lumping them into “with atmospheric sound” gets you a generic bed.

1 — Room tone

The sound of the space itself, with nothing happening in it. This is the layer that sells the location and the one people most often forget. “A large empty concrete stairwell, considerable natural reverb, distant traffic bleeding through glass” does enormous work.

2 — Action sound

Sound tied to something visible in frame. This is where the model is genuinely impressive, and where specificity pays: “the dry scuff of a leather sole on concrete, the heavy rustle of a wool coat as she turns”. Naming the material gives you the right sound, exactly as naming the fibre gives you the right drape.

3 — Voice

Dialogue or breath. The most powerful layer and the one that most rewards direction, covered next.

Writing Dialogue That Arrives Usable

Three rules, and the third is the one people learn the hard way.

Put the words in quotation marks

Quote the line exactly, attributed to a described speaker. Unquoted, the model tends to paraphrase or to treat your line as a description of the subject matter rather than the words to be said.

Direct the delivery separately from the line

An undirected line arrives in a flat, announcer-ish register — the audio equivalent of frontal lighting. Tone, pace and volume are separate instructions and each lands: “spoken quietly, almost under her breath, unhurried, as if finishing a thought rather than starting one”.

Write far less than you think

This is the rule that catches everyone. A few seconds of screen time is one short sentence. Write a paragraph and the model either rushes it into unusable garble or clips it mid-word at the end of the clip. Count it out loud against the clip length before you generate.

A Worked Prompt

Picture and sound directed together, one short line, clip kept short:

Note what is specified: one movement, one line, one held beat. Six seconds cannot hold more than that, and prompts fail far more often from overstuffing than from underwriting.

Start From a Frame, Not a Description

Whenever the subject matters — and in campaign work it always does — hand the model an approved still as the clip's first frame rather than describing who is in it.

A start frame anchors casting, wardrobe, palette and light in a way no amount of written description can, and it is the only practical route to the same subject appearing across several clips in one campaign. It also means the still and the film share a single source of truth, which is how continuity is supposed to work.

Describe only what changes: the movement, the camera, the sound. The frame has already answered everything else.

Five Ways Clips Become Unusable

Generated ambience under a licensed track

The most common and most avoidable. If the piece is scored, generate silent. Two competing sound worlds cannot be fixed in the edit.

Dialogue that runs past the clip

Written for length rather than duration. Say the line aloud, time it, then cut it roughly in half.

A camera move that fights the subject

Dolly in while the subject turns and something usually breaks — most often the face. Move the camera or move the subject, rarely both inside a few seconds.

Sustained close-up on a face

Identity drift is a function of duration. A long hold on a face is the single hardest ask in AI video. Shorten it, or design around it.

No cut point

Clips that start and end mid-movement are hard to cut together. End on a held beat where you can. It gives the editor a place to land.

Frequently asked

Does Veo 3.1 generate audio as well as video?

Yes — native audio alongside the picture, including ambience, effects tied to on-screen action, and spoken dialogue. Sound becomes something you direct rather than add in post, and leaving it unspecified means the model chooses for you.

How do you write dialogue in a prompt?

Put the words in quotation marks, attribute them to a described speaker, and direct the delivery separately from the line. Keep it short: a few seconds of screen time is one short sentence, not a paragraph.

How long should a clip be?

Four to eight seconds is the practical working range. Identity and audio quality both degrade with duration. Cut two short clips together rather than asking for one long take.

Can you turn the audio off?

Yes, and you should whenever the piece is cut to a music bed. Generated ambience fighting a licensed track is one of the most common reasons a good clip becomes unusable.

Should you start from an image?

Whenever the subject matters, yes. An approved still as the first frame anchors casting, wardrobe, palette and light far more reliably than description, and it is the only practical way to hold one subject across several clips.