Veo 3 prompt guide: what our counted runs taught us

Oct 29, 2026

Most Veo 3 prompt guides are written from other people's clips. This one is written from our own generation receipts. On Oct 29, 2026 we spent one full day of Google Flow credits — 50, the daily allowance on our account — on exactly two prompts. Four generations later we had two finished clips, and every attempt, including the ones we discarded, is logged on the entry pages. What follows is what those runs taught us about how to prompt Veo 3, with both tested examples embedded in full. It is a small sample, and we say so wherever a lesson rests on a single run.

1. One paragraph of plain prose

Both prompts we ran are a single paragraph pasted into Google Flow — no structure tags, no code, no parameter syntax. Here is the first one, exactly as generated:

A cozy café interior in warm morning light. A young woman with curly hair sits across from an older man at a small wooden table, two coffee cups between them. She smiles and says, in a cheerful tone: 'I finally launched my website today!' He raises his cup and replies warmly: 'To new beginnings.' Soft café ambience, gentle background chatter, shallow depth of field, cinematic natural lighting, photorealistic, 35mm film look.

Read its order, because the order is the method: what we see (the café, the two people), what happens (a toast), what is said (the two lines), how it is shot (depth of field, lighting, film look), what it sounds like (ambience, chatter). A Veo 3 prompt is a tiny shooting script written as prose. When a generation disappoints, the fix is usually a missing sentence in that order, not more adjectives.

2. Dialogue: quotes, a named speaker, a tone

The spoken lines sit in quotes inside the prose, each attached to a person and a tone — "in a cheerful tone" for her, "replies warmly" for him. The result is the single most encouraging data point in our runs: the first generation, on Veo 3.1 Lite — the cheapest tier — already had both lines spoken by the right people, in the right order. Nothing about the dialogue needed a second attempt. If two people speak in your scene, give each line its own clause with its speaker named; do not make the model guess who says what.

3. Sound is part of the prompt

Veo generates audio with the picture, and our second run was built around that. The prompt, exactly as generated:

Extreme macro close-up: a crystal-clear glass strawberry on a dark slate board. A chef's knife slowly slices through it with a crisp, satisfying glass-crack ASMR sound, the slice falling away to reveal sparkling glass segments inside. Studio lighting with sharp reflections and refractions, black background, ultra-detailed, photorealistic, no music, only the natural cutting sounds.

Two sound moves are worth copying. First, the wanted sound is described as precisely as the picture — "a crisp, satisfying glass-crack ASMR sound", not "with sound effects". Second, the prompt ends with a negative: "no music, only the natural cutting sounds". The café prompt does the same job positively, naming the ambience it wants. Write the soundtrack as plainly as the cinematography, including what should not be heard.

4. Draft on Lite, finish on Fast

The café run shows the spending pattern we now use by default. Generation 1 ran on Veo 3.1 Lite for 10 credits and proved the prompt: scene, staging and speech all correct. Its one flaw was flat café lighting. Generation 2 ran on Veo 3.1 Fast for 20 credits with the same text, and fixed exactly that — warmer light, and the cup-raise on "To new beginnings" reads clearly. Total: 2 generations, 30 credits, one finished clip.

Spend cheap generations proving the prompt, and expensive ones polishing the keeper. And decide the single flaw you are fixing before you regenerate — see the next section.

5. Fix one flaw per generation

The ASMR run needed no model upgrade, because its flaw was timing, not fidelity. Generation 1 (Lite, 10 credits) was usable, but the glass cracked only as the slice separated. The shot the prompt asks for is the crack while the knife is still pressing in — so generation 2 (Lite again, 10 credits) targeted that single timing flaw, and the shipped final cracks at the right moment. Total: 2 generations, 20 credits.

The discipline generalises: name the flaw in one sentence before regenerating. If you cannot name it, you are not iterating — you are gambling credits.

6. Pick the aspect for the subject

The café scene ran at 16:9 — two people, a table, a room that needs width. The glass-fruit shot ran at 9:16 vertical — a single object, shot for a phone screen. Both of our runs are 8-second clips. Aspect is a prompt-adjacent decision, made before generating, from the subject — not a default you accept and crop later.

7. Verify like a skeptic

One honest caveat, copied from the ASMR entry's own log: the glass-crack ASMR sound could not be fully verified in remote playback. There is no speech and no music — consistent with what the prompt requests — but the fine crackle should be checked on headphones before anyone relies on it. A generation that looks right can still be wrong in the parts you cannot inspect from where you sit. Log what you could not verify, on the entry, in public. That is what the "What broke" section on every entry on this site is for.

The 50-credit day, totalled

Two finished, published clips for one day of free credits — because each prompt was proved cheap before it was polished. As the daily credits refresh, the tested Veo set grows; you can watch it happen on the Veo tag page, where every entry carries its generation count. For the general method behind our prompts — structure, banned lists, beat grids — see how to write a motion prompt that actually reproduces.

Method note: this guide rests on two prompts and four generations, run on Oct 29, 2026 in Google Flow (Veo 3.1 Lite and Fast, 50 credits total). Lessons drawn from a single run are framed as method, not measurement. As the tested Veo set grows we will update this guide and say what changed.