Veo 3.1 responds to structure far more reliably than to length. A prompt that names the shot, the subject, the action, the setting, and the mood — in that order — produces a controllable result; the same idea written as one long descriptive sentence does not. This guide covers the formula, the audio syntax that makes Veo 3.1 different from earlier models, how to use Fast mode without wasting renders, and the negatives that fix the most common failures.
First, the model landscape
Veo 3.1 is the current Veo model. It generates 720p or 1080p video with native synchronised audio — dialogue, sound effects, and ambience produced together with the picture rather than added afterwards. It ships in three variants:
| Variant | What it is for |
|---|---|
| Veo 3.1 | The full model. Cleanest physics, most stable camera work. Use for the render you keep. |
| Veo 3.1 Fast | Cheap, quick iteration. Use to find out whether a prompt works before committing. |
| Veo 3.1 Lite | Free tier, available on any Google account. Expect more re-rolls on dialogue and fine detail. |
There is no Veo 4. At Google I/O 2026 the next-generation video model was announced as Gemini Omni Flash — a separate line that accepts image, audio, video, and text as input and caps output at ten seconds — not a Veo 4 release. If you have landed on a page promising “Veo 4 prompts,” it is describing a model that does not exist.
The five-part formula
Every reliable Veo 3.1 prompt has the same skeleton:
[Shot composition] + [Subject details] + [Action] + [Setting] + [Aesthetic / mood]Order matters. Veo weights earlier tokens more heavily, so anything at the end is the first thing it compromises on. That is why shot type and subject go first and mood goes last — if the model has to drop something, you want it dropping “wistful” rather than “close-up.”
Underspecified:
A woman drinking coffee, cinematic.Five-part:
Medium close-up, shallow depth of field. A woman in her forties with
grey-streaked hair sits at a window table, both hands around a ceramic mug.
She lifts it, breathes in, and takes a slow sip without looking away from
the window. A quiet cafe on an overcast morning, condensation on the glass.
Muted palette, soft diffused light, unhurried.The second version is not better because it is longer. It is better because every clause removes a decision Veo would otherwise make differently on each generation. The first prompt leaves roughly a dozen variables unstated; the second leaves two or three.
Writing the audio
Native audio is the main reason to prompt Veo 3.1 differently from an older text-to-video model. There are three kinds of audio direction and they behave differently.
Dialogue goes in quotation marks after a says-style verb, with the tone stated outside the quote:
She turns to the camera and says, in a flat deadpan: "That's the third
time this week."Keep spoken lines to roughly sixteen words or fewer for a short clip. Longer lines force Veo to compress the delivery, and compressed delivery is where lip-sync drifts. If you need more words, you need more shots.
Sound effects are most reliable when tied to a visible action, because the model then has something to synchronise against:
The lid clicks shut and the latch snaps into place.Ambience sets the floor of the mix and should be named explicitly, or Veo will invent it:
Room tone with a faint refrigerator hum underneath, no music.That last clause matters. Left unspecified, Veo 3.1 frequently adds a generic music bed. Naming the ambience and excluding music is more reliable than either alone.
Fast mode: what it is actually for
Veo 3.1 Fast is not a lower-quality model you settle for — it is an iteration tool, and using it correctly is the single biggest saving available in a Veo workflow.
The pattern that works:
- Draft the prompt using the five-part formula.
- Run it on Fast. You are checking structure only: is the subject right, does the camera move do what you meant, does the line fit the clip?
- Fix the structural problem. Do not fix lighting or grain yet — Fast will not represent those faithfully.
- Once the structure holds, run the identical prompt on full Veo 3.1.
The mistake to avoid is iterating on aesthetics in Fast mode. Fine detail, subtle lighting, and material texture are exactly what the fast variant approximates, so polishing them there tells you nothing about the final render. Iterate structure in Fast; judge finish in full.
Negative prompts
Negatives work in Veo 3.1, but they work best aimed at the model’s own defaults rather than at arbitrary objects. Veo reliably adds a small set of things when you do not say otherwise:
| Default it adds | Negative that removes it |
|---|---|
| Burned-in subtitles | no on-screen captions, no text overlay |
| A generic music bed | no background music |
| A slow creeping zoom | no zoom, locked-off camera |
| Extra people in frame | no other people in frame |
| Over-saturated grade | natural colour, no heavy grade |
Phrase negatives as descriptions of absence rather than as commands. “No background music” outperforms “don’t add music,” because the prompt is read as a description of the finished clip, not as an instruction to an assistant.
One caution: negatives cannot rescue a vague positive. If the subject is underspecified, five negatives will not stabilise it — tighten the subject description first, then use negatives to trim the defaults.
Common failures and what causes them
The clip feels rushed. You asked for too many beats. One continuous action per generation; if you need a setup and a payoff, that is two prompts.
The subject changes between generations. The subject description is doing less work than you think. “A woman” leaves everything open; “a woman in her forties with grey-streaked hair, in a navy wool coat” pins it down. For a character you need repeatedly, keep the description verbatim across prompts.
Lip-sync drifts. The line is too long, or the tone was written inside the quotation marks and got spoken aloud. Shorten the line and move the tone outside the quote.
The camera does something you did not ask for. You did not name a camera state. “Locked-off” and “static” are instructions too — an unnamed camera is an invitation to drift.
Text on a sign or product comes out garbled. Veo 3.1 is unreliable at rendering specific lettering. Frame the shot so text is small, angled, or partly out of frame rather than fighting it.
Where to go next
For the structure itself in more depth, see the prompt formula and structure guide. For camera vocabulary Veo reliably recognises, see the camera movement guide. For batching and versioning prompts, see the JSON prompt format guide, and for fixing a specific broken output, the negative prompt troubleshooting guide.
If you would rather assemble the five parts through a form than write them out by hand, the free Veo prompt builder does exactly that and outputs both Text and JSON. Working examples by use case live in the Veo prompt library.