Production Safety

Prompt Safety Guide

US Video API uses a commercial foundation model with safety review. That is a feature of production infrastructure, not a bug. This guide shows what tends to get blocked, how to rewrite prompts, and how to design a workflow that still ships on time.

Why prompts get blocked

Commercial video models do not only inspect the words you typed. They also assess the likely output. A prompt can look harmless and still be rejected if the model predicts spoken dialogue, sensitive audio, real-person likeness, or ambiguous brand/IP output.

Most common production blockers: implied speech, celebrity or real-person references, copyrighted characters, minors, explicit violence, brand/logo requests, and prompts that ask for audio-specific output.

What to avoid by default

Higher-risk prompt patterns
  • “A spokesperson explains...”
  • “Make it sound like...”
  • Celebrity, politician, or influencer likeness
  • Brand mascot or copyrighted character references
  • “Exact ad for Brand X” or visible logo instructions
  • Anything that depends on dialogue to make sense
Lower-risk prompt patterns
  • Silent visual-only scenes
  • Product shots, environments, motion, lighting
  • Abstract workflow or dashboard visuals
  • Workspace details with no people
  • Clear camera direction and material description
  • Negative prompts excluding speech, text, and logos

Rewrite patterns that work

The safest production strategy is to keep prompts literal, commercial, and visual-first.

Pattern 1: remove dialogue

Instead of:
"A founder speaks directly to camera about why this software is better"

Rewrite to:
"A silent visual-only enterprise promo, elegant office environment, dashboard screens, subtle dolly camera, no people speaking, no readable text"

Pattern 2: remove real people

Instead of:
"A creator like [real person] records a tutorial at a desk"

Rewrite to:
"A premium creator workspace sequence, camera gear close-up, keyboard glow, microphone on clean desk, soft daylight, no people, no speech"

Pattern 3: remove direct brand/IP references

Instead of:
"A luxury ad for Brand X with their exact bottle"

Rewrite to:
"A premium skincare product showcase, single hero bottle on polished stone pedestal, soft mist, luxury studio lighting, realistic packaging materials, no logos"

Build fallback logic into the product

Do not assume every first-pass prompt will clear moderation. Production systems need graceful degradation.

Recommended fallback chain:
  1. Try the original prompt.
  2. If blocked, remove dialogue and audio cues.
  3. If still blocked, remove real people, brands, and IP references.
  4. If still blocked, switch to a silent visual-only template for the same use case.

For high-volume use, maintain a prompt library with pre-approved product, enterprise, creator, and abstract workflow templates.

What to tell your users

You should disclose that the model has safety review, but present it as a commercial boundary, not a platform defect.

Recommended wording: “Some requests may be blocked by model safety review. For best results, use commercial, visual-first prompts and avoid real-person likenesses, copyrighted characters, and dialogue-heavy scenes.”

That sets the right expectation while still sounding like production infrastructure, not consumer-grade unpredictability.

Examples

Safe enterprise example

{
  "prompt": "A premium enterprise technology promo with no people and no speech, elegant control room environment with large analytics displays showing abstract video workflow activity, polished commercial frames animating across clean monitors, subtle dolly camera movement, restrained deep blue and neutral palette, no readable text, no logos, silent visual-only scene",
  "negative_prompt": "people, faces, hands, speech, captions, readable text, clutter, cyberpunk, low realism"
}

Safe creator example

{
  "prompt": "A premium vertical social video with no people and no speech, camera lens close-up, microphone on a clean desk, editing keyboard glow, soft daylight mixed with warm practical lights, subtle handheld-style push-ins, polished b-roll aesthetic for Shorts, Reels, and TikTok, no readable text, silent visual-only sequence",
  "negative_prompt": "people, faces, hands, speech, captions, logos, messy room, distorted gear, low realism"
}
Next steps: pair this guide with the API documentation for endpoint behavior and the enterprise guide if your team needs shared access, billing ownership, and multi-user controls.