Prompting11 min read

Veo-Style "No/Don't" Prompts Without Negative Prompts: 12 Cleaner Rewrite Patterns

Learn 12 clean rewrite patterns to avoid “no/don’t” prompt backfires—using positive specs + undesired tokens for more predictable AI video results.

On this page

TL;DR

If “no/don’t” lines keep summoning the thing you’re trying to avoid, stop writing negatives as sentences. Instead:

  1. Over‑specify what you do want (subject → action → scene → camera → lighting/style → optional audio).
  2. Add a compact undesired tokens list (nouns like watermark, logo, border).
  3. Iterate by changing one lever at a time.

This aligns with Google’s video prompt guidance: use correct keywords/structure, break the idea into components, and use token-style lists for unwanted items rather than “don’t” phrasing. (https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/video-gen-prompt-guide)

Key takeaways

Why “no/don’t” often backfires in video prompts

A negative sentence still contains the visual concept.

  • “No logo” includes the token logo.
  • “Don’t add text” includes text.
  • “Don’t be shaky” doesn’t define the allowed camera behavior.

Google’s prompt guide emphasizes that prompt structure and correct keywords help the model create what you want—and it highlights token-style approaches for steering away from unwanted items. (https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/video-gen-prompt-guide)

The practical fix is simple:

  1. Crowd out defaults with precise positives.
  2. Name intruders as short undesired tokens.

The reusable template: Positive Spec + Undesired Tokens

Google recommends breaking an idea into key components and notes you don’t need every element in every prompt—use the elements that buy you control. (https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/video-gen-prompt-guide)

Copy‑paste prompt skeleton

Use this for text-to-video or image-to-video.

Subject / scene (who/what + where):

Motion / action (what changes over time):

Camera (what the viewer experiences):

  • Shot size + movement you allow (e.g., “locked-off tripod,” “slow dolly-in”).

Style / lighting / mood:

  • Lighting behavior and the look you want.

(Optional) Audio:

  • Simple, diegetic instructions (dialogue/SFX/music). If your generator supports native audio, define it plainly.

Undesired tokens (short noun list):

  • text, watermark, logo, subtitles, UI, border, frame, flicker, glitch

Worked example (before/after)

This is the pattern you should A/B test first because it replaces vague prohibitions with concrete positives.

Before (negative sentences):

“A clean product video of a skincare bottle on a marble counter. Don’t show any logos or text. No weird hands. Don’t zoom randomly. No flicker.”

After (positive spec + undesired tokens):

  • Subject/scene: Frosted glass skincare bottle, centered on a white marble counter in a bright minimalist bathroom. Blank white label with no markings.
  • Motion: Bottle stays still; soft morning light shifts subtly; a single condensation bead slides down the glass.
  • Camera: Locked-off tripod, medium close-up.
  • Style: Clean commercial realism, soft diffuse lighting, neutral color grade.
  • Undesired tokens: text, watermark, logo, brand mark, subtitles, UI, border, frame, flicker, glitch, extra fingers, cropped hands, rapid zoom

What changed (mechanically):

  • “Don’t show text” → “blank label” (a positive surface spec).
  • “Don’t zoom” → “locked-off tripod” (allowed camera behavior).
  • “No weird hands” → specific failure modes as tokens.

12 cleaner rewrite patterns (no negative prompts required)

Use 1–3 patterns per iteration. More than that and you won’t know what fixed (or broke) the result.

1) Replace “no text” with blank / unbranded surfaces

Before: “No text or labels.”

After: “Plain blank label; smooth unmarked packaging surfaces.”

Undesired tokens: text, typography, watermark, logo

2) Convert “don’t show X” into “show Y instead”

Before: “Don’t show a crowd.”

After: “Single person alone; empty background; no bystanders in frame.”

Undesired tokens: crowd, pedestrians, passersby

3) Turn vague negatives into a concrete undesired tokens list

Google’s guide supports token-style steering away from unwanted items. (https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/video-gen-prompt-guide)

Before: “Don’t add weird artifacts.”

After: “Undesired tokens: flicker, glitch, banding, warping, compression artifacts.”

4) Stop “default clutter” by specifying a set

Before: “Talking head video, no messy room.”

After: “Plain matte off-white wall; no decor; soft shadow falloff.”

Undesired tokens: posters, shelves, frames, clutter, plants

5) Fix borders/overlays by describing full-bleed framing

Before: “No black bars or borders.”

After: “Full-bleed edge-to-edge frame; clean corners; no overlay.”

Undesired tokens: border, frame, vignette, letterbox, pillarbox, UI

6) Prevent surprise edits with single continuous shot

Before: “Don’t cut away.”

After: “Single continuous shot from start to end; no transitions.”

Undesired tokens: cut, transition, montage, jump cut

7) Stop random zooms by defining allowed movement

Before: “Don’t zoom in and out.”

After: “Camera fixed; if any movement occurs, it’s a slow subtle dolly-in over the full clip.”

Undesired tokens: snap zoom, rapid zoom, whip pan, shaky cam

8) Fix hands by forcing visibility + simplicity

Before: “No extra fingers.”

After: “Two hands only; fully visible from wrist to fingertips; one simple action: place the object down and release.”

Undesired tokens: extra fingers, fused fingers, cropped hands, distorted hands

9) Remove unwanted objects by controlling the props inventory

Before: “No random objects on the table.”

After: “Only three items in frame: the mug, one spoon, one folded napkin; clean tabletop.”

Undesired tokens: phones, keys, wrappers, clutter

10) Reduce flicker by specifying lighting behavior

Before: “Don’t flicker.”

After: “Stable continuous lighting; consistent exposure; soft diffuse key light.”

Undesired tokens: flicker, strobe, exposure pumping

11) Prevent style drift with a tight style anchor

EachLabs warns that too few details leads to guessing and too many can confuse—so anchor the style with a small set of compatible descriptors. (https://www.eachlabs.ai/blog/image-to-video-prompt-guide-best-practices-for-realistic-results)

Before: “Make it cinematic, not cartoonish, not anime.”

After: “Photoreal documentary-style color grade; natural skin texture; real lens behavior.”

Undesired tokens: cartoon, anime, illustration, 3d render

12) When you must control on-screen text: keep it short + spatial

Before: “Show this full sentence on screen: ‘Limited time offer…’”

After: “On-screen text: ‘SALE’ (4 letters), top-left on a solid banner; wide framing so the word is fully readable.”

Undesired tokens: small text, paragraphs, scrolling text, distorted typography

Quick A/B test workflow (so you don’t prompt-churn)

Google’s guide explains modifying prompts to produce different results/effects; do that systematically, not emotionally. (https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/video-gen-prompt-guide)

10-minute loop

  1. Save your base prompt as v1 (Subject + Action + Camera + Style).
  2. Pick one symptom (e.g., “random logo appears”).
  3. Apply one pattern (e.g., Pattern #1 “blank label”).
  4. Generate variants where only one variable changes:
    • A: positive spec change only
    • B: undesired tokens change only
    • C: both
  5. Keep the best, then tackle the next symptom.

What to change first (table)

Symptom you see First change (positive spec) Second change (tokens)
Random logos/text “blank/unbranded surface” text, watermark, logo
Random zooms “locked-off tripod / allowed movement” snap zoom, whip pan
Flicker/exposure pumping “stable continuous lighting” flicker, strobe, exposure pumping
Hand artifacts “two hands, fully visible, one action” anatomy tokens (extra fingers, etc.)

If you want to run this kind of structured batch testing at scale, Veo3Gen offers a developer API so you can generate variants programmatically and keep prompt versions consistent across campaigns.

Where Veo3Gen fits (practical selection)

If you’re applying these rewrite patterns, you’ll typically need two things: throughput (more iterations) and control (fewer surprises).

Veo3Gen is an affordable way to access Google’s Veo 3.1 video models without Google’s enterprise pricing. It offers three modes—Veo 3.1 Fast (quick, great default), Veo 3.1 Quality (max fidelity), and Veo 3.1 Lite (cheapest, preview). It supports text-to-video and image-to-video, plus first-and-last-frame control on Veo 3.1. Supported resolutions include 720p, 1080p, and 4K (4K on Fast/Quality), with 16:9 and 9:16 aspect ratios. Generations include native, synchronized audio (dialogue, SFX, music) in a single pass. New users get free credits to start, and pricing is pay-as-you-go credits plus optional monthly plans; purchased credits do not expire.

Mid-article CTA: If your current workflow is “prompt → regret → reprompt,” try running the A/B loop above inside Veo3Gen using Veo 3.1 Fast as your default iteration mode, then switch to Quality once your prompt is stable.

Troubleshooting (when the unwanted thing still shows up)

Tighten subject specificity

Google notes subject specificity helps avoid generic outputs. If your subject is “a person,” the model will invent defaults. Make it concrete: wardrobe, location, framing. (https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/video-gen-prompt-guide)

Remove competing instructions

EachLabs warns that too many details can confuse the model. If you’re fighting artifacts, reduce to one location + one action + one camera behavior. (https://www.eachlabs.ai/blog/image-to-video-prompt-guide-best-practices-for-realistic-results)

Move constraints into camera language

“Locked-off tripod” beats “don’t be shaky.” “Single continuous shot” beats “don’t cut.” (Also consistent with the “camera language” emphasis seen in other prompt guides.) (https://kling.ai/blog/kling-ai-prompt-guide)

Safety blocks

Google states safety filters apply and prompts violating responsible AI guidelines can be blocked. If you’re blocked repeatedly, rewrite the concept; don’t keep rephrasing the same disallowed intent. (https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/video-gen-prompt-guide)

Checklist

  • Write the shot as Subject/scene → Motion → Camera → Style → (Optional) Audio.
  • Replace every “no/don’t” with a positive alternative (e.g., “blank label,” “plain wall,” “locked-off tripod”).
  • Add a compact undesired tokens list (short, noun-like terms you can point to).
  • Constrain editing: single continuous shot unless you explicitly want cuts.
  • Constrain camera: state allowed movement (or none).
  • If hands appear: require two hands, fully visible, one simple action.
  • If text must appear: keep it short and specify placement + framing.
  • Iterate by changing one pattern at a time.

FAQ

### How do I remove unwanted objects in AI video without negative prompts?

Specify a simple set (background + props inventory), then list common intruders as undesired tokens (e.g., clutter, posters, signage).

### How do I stop random text, watermarks, or logos from appearing?

Use a positive surface spec like “blank label / unmarked surface,” then add text, watermark, logo, subtitles as undesired tokens. Avoid writing “no text” as a sentence. (https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/video-gen-prompt-guide)

### How do I prevent random zooms and shaky camera moves?

Define what’s allowed: “locked-off tripod” or “slow dolly-in only,” then add snap zoom, whip pan, shaky cam to undesired tokens.

### How do I reduce flicker in AI-generated video?

Specify stable lighting/exposure behavior (“stable continuous lighting; consistent exposure”), then add flicker, strobe, exposure pumping as undesired tokens.

### How do I get readable on-screen text in AI video?

Keep the text short and spatial: a few characters, clear placement, and framing that keeps it legible. (FlexClip’s structure reminds you to define scene + camera; treat text as part of that.) (https://help.flexclip.com/en/articles/10326783-how-to-write-effective-text-prompts-to-generate-ai-videos)

Create cleaner videos faster with Veo3Gen

These 12 patterns work best when you can generate, compare, and lock a “house prompt” quickly.

Veo3Gen gives you access to Google’s Veo 3.1 video models with text-to-video, image-to-video, first-and-last-frame control, and native synchronized audio in a single pass—so your “clean visual prompt” can include simple dialogue/SFX/music without a separate audio step.

Closing CTA: Start with your ugliest “no/don’t” prompt, rewrite it using Pattern #1 + Pattern #7, and run a small A/B batch in Veo3Gen (Fast for iteration, then Quality once it’s locked). Then save the winning version as your reusable template for the next 10 videos.

Start creating with Veo3Gen

Veo3Gen gives you affordable Veo 3.1 video generation with native audio, up to 4K, and credits that never expire — with free credits to start.

Limited Time Offer

Try Veo 3 & Veo 3 API for Free

Experience cinematic AI video generation at the industry's lowest price point. No credit card required to start.

Get Started FreeView Cheapest Pricing
← Back to all posts