Veo Flow Prompting

Veo 3.1 / Gemini Omni Flash prompting

Field-tested rules for Google Flow (labs.google/fx/tools/flow). Full source-cited background: docs/google-flow-veo-playbook.md. UI map and credit costs: docs/flow-ui-map.md. Filters and policy: docs/content-filters-and-policy.md.

Claim tags used throughout: [FACT] = stated by a Google primary source · [ESTIMATE] = practitioner guidance · [UNVERIFIED] / [CONTESTED] = confirm in-product before relying on it. Never invent model behaviour; if you are unsure, tag it.

Which model is which. "Omni Flash" is Gemini Omni Flash, Google's any-input (text/image/audio/video) → video model with native audio. It is not a Gemini text LLM and there is no product called "Veo 3 Flash". Veo 3.1 (Lite / Fast / Quality) is the cinematic model. Both are selectable in Flow's model dropdown. [FACT]

1. The per-clip prompt block

Field order, every clip: LOOK → CAMERA → ACTION → FIDELITY → DIALOGUE → AUDIO → EXCLUSIONS (+ optional POST NOTE).

Google's own five-part formula for Veo 3.1 is [Cinematography] + [Subject] + [Action] + [Context] + [Style & Ambiance] [FACT]. The block above is that formula plus the three lines production keeps needing (fidelity, audio direction, exclusions).

LOOK:       <subject + wardrobe + setting + light, restated in full>
CAMERA:     <exactly ONE move>
ACTION:     <max 2 beats>
FIDELITY:   Keep natural, realistic body proportions. Do not slim or idealize.
DIALOGUE:   He says, "<one or two short lines, in quotes, attributed>"
AUDIO:      Ambient noise: <...>. SFX: <...>.
EXCLUSIONS: a clean frame with no subtitles, no captions, no on-screen text,
            no lower-thirds, no logos, no watermark.

Every clip prompt must be self-contained. The generator never sees your other clips. Repeat wardrobe, setting and identity in every clip. A "base card" at the top of your script document is human reference only — it is not merged into anything.

Complexity budget per clip (production-learned; overloaded clips render wrong):

• Max 2 action beats. More than that → split into two clips.
• Exactly one camera move [FACT]. Vague or contradictory camera cues make the model default to near-static. For a locked frame, say static shot, camera completely still [ESTIMATE] — omitting the camera line does not give you a locked frame.
• No invisible micro-props (pressure sensors, hidden latches). Imply them through timing instead: "the moment it leaves the cushion, the alarm sounds".
• State changes only at the very end of the clip — never a mid-clip lighting transition inside the LOOK line.
• In close-ups keep one character in focus; others soft-blurred or off-screen with off-screen dialogue.

Length: ~100–200 words per prompt [ESTIMATE]. Beats of 4–8s. Omni Flash goes to 10s; Veo 3.1 caps at 8s. 1080p/4K require an 8s clip; Extend caps at 720p [FACT].

Lens to feel: 16mm expands space · 35mm natural · 50mm intimate · 85mm compresses for intensity · wide-angle for establishers [ESTIMATE].

Timestamp prompting packs several shots into one generation [FACT]:

[00:00-00:03] Medium close-up, slow push-in: <shot>.
[00:03-00:07] Static medium shot: <shot>.
[00:07-00:10] Slight low angle, locked frame: <shot>.

Anchor performance to words: "firm flipping gesture on 'flip that'" works; "expressive gestures" does not.

2. Dialogue, language and silence

• Write directions in English (best visual control) and dialogue in the target language. This pattern is production-verified.
• Non-English dialogue must keep its accents — nao → não, e → é, missao → missão. Accents drive pronunciation; never ASCII-strip a dialogue line. Spell numbers out in words.
• Speech is attributed to a named speaker, and kept to 1–2 short lines per clip. On quotation marks the sources conflict [CONTESTED]: Google's Veo prompt guide shows quoted dialogue (A woman says, "We have to leave now."), while separate Google guidance says to use a colon after the speaker's action and avoid quotation marks, because quotes push the model toward rendering the text on screen. Field testing sided with the colon form, and it composes with the caption bug below — so default to Man says: Bom dia! and keep the quoted form as a fallback if speech fails to trigger at all. Full detail: dialogue-and-voice.
• No ellipsis and no mid-sentence periods in a dialogue line — they stop the voice and the lip-sync dead. Use commas; punctuate only at the very end.
• Name the language on every spoken line (and every background murmur). Omitting it causes audio failures and random-language output.
• Regional accent control: tag every line, e.g. DIALOGUE (PT-BR, carioca accent - coda "S" softened to "sh"): "...". Accent adherence is unstable [UNVERIFIED]; fallback is to record or TTS in post.
• Silent clip recipe: no quoted speech, no speech cues at all, positively specify the ambient audio, then add no dialogue, no speech, silent. Speech cues are the primary trigger for both spoken audio and burned-in captions [FACT]. Flow also has an explicit toggle: View Settings → Return silent videos: On. Never mention the mouth, lips or speaking in a silent clip — "his mouth opens slightly as if about to speak" reliably generates actual speech. Emotion goes through eyebrows, jaw and eyes.
• Audio is always generated and there is no mute flag [FACT] — so direct it per clip. If only one clip should carry crowd noise, say so and keep the others quiet.

3. The gotchas that cost credits

• Burned-in junk captions. Veo burns garbled subtitles into clips even when told not to [FACT]. Google's documented mechanism is negative prompting written as positive description — describe the clean frame you want, not "no X". End every clip with the EXCLUSIONS tail above. It is effective but not foolproof; inspect every render.
• On-screen text is garbled, including short words and accented words. Generate the plate clean and add real text in post, or use Flow's Type Overlays tool. Same for end cards.
• Branded real locations get approximated. Describe them generically ("a grand white Belle Époque beachfront hotel"), optionally add a reference photo of the place, and never ask for the brand name or logo on screen — exclude it.
• SynthID watermark is always on, no opt-out [FACT]. Label your output as AI-generated where the platform requires it.
• PG / stylized action renders better and avoids refusals: no explicit weapons, no blood, no real explosions. Alarms, chases and heists are fine.
• Aspect ratio: 9:16 for Reels/Shorts, 16:9 for YouTube/TV. If the UI already sets the ratio, do not also write "Vertical 9:16" in the prompt text.
• Credits are spent per generation, and there are no free re-rolls for successful generations [FACT]. Validate identity with one cheap, calm close-up before committing to a whole sequence. (You are not charged for failed generations [FACT] — but see the NPOV note below: a filtered generation in Flow's UI has been observed to charge anyway.)

4. When the content filter blocks you

Flow returns a single generic message — "This prompt might violate our policy about generating harmful content" — for completely different underlying causes. Do not guess which one you hit. Full detail, including how to read the real reason out of the API response, is in docs/content-filters-and-policy.md.

The one finding that changes how you write prompts:

NPOV — the news framing itself is the trigger, not the faces. Verified by inspecting the generation-status API: a blocked generation returns failureReasons: ["NPOV"] (Neutral Point Of View). Framing a clip as a news broadcast — anchor, newsroom, "delivering the news", and even a news-coded set (channel logo wall, world map panel, anchor desk) — fails, no matter how neutral the content. The same real events told as a casual story or anecdote pass, with the same people, the same dry wit and the same facts.

The fix: generate "two people sharing absurd stories to camera, dry deadpan amusement; not a news broadcast" over a generic modern studio, and composite the newscast look (logo, chyrons, ident) 100% in post. The model sees storytelling; the audience sees a newscast.

Related: naming a real celebrity in dialogue may trip a separate prominent-people filter. Prefer the no-name variant by default ("a famous player", "a European airport") so you never gamble credits on it.

5. Before you deliver a multi-clip script

Run the continuity protocol and produce a regeneration ledger — both live in skills/multiclip-continuity. Skipping this is the single most expensive mistake in multi-clip work: props teleport, wardrobe changes between clips, and characters act on facts they were never told.

Identity, avatars, reference photos and consent: skills/identity-and-likeness.

6. Honesty rules

• Big claims in a script are framed as vision ("what if…"), not fact.
• A statistic needs a source in hand before it appears as fact in a script.
• Mark every unconfirmed model behaviour [UNVERIFIED] or [CONTESTED] and say so out loud — model behaviour and pricing move fast, and a stale "fact" in a prompt playbook burns real money.