# Image Generation Guidelines for AutoShortsBot

Current strategy: **TrueTraceShorts — Digital Red Flags / Scam Self-Defense for Normal People**.

Core rule:

> One screen. One red flag. One safer move.

## Default Boundary

For scam-screen content, deterministic renderer-owned mockups are usually better than free-form AI images.

Deterministic renderer should own:

- SMS / delivery text screens
- invoices and payment routes
- login pages
- marketplace chats
- popup warnings
- screenshot checklist cards
- any factual text, URLs, amounts, warning labels, or safe-move instructions

Use AI image generation for:

- premium atmosphere
- realistic phone/laptop/desk environment
- text-free background or styleframe texture
- cinematic first-frame mood when the critical scam screen is still controlled

## Prompt Template for AI Atmosphere/Styleframes

```text
AutoShortsBot premium keyframe, vertical 9:16 social-video frame.
Series: Digital Red Flags / Scam Self-Defense for normal people.
Scene role: <hook / visible trap / safer move / takeaway>.
Visible real-world problem: <what is concretely broken in the first second>.
Subject and environment: <phone/laptop/desk/package/marketplace scene>.
Composition: <camera angle, focal point, negative space, safe margins>.
Lighting/style: realistic premium social video, not abstract AI art.
Renderer-owned screen: leave area for deterministic mockup if text/data matters.
Forbidden: logos, brands, QR codes, real bank data, phone numbers, real domains, people/faces/hands unless explicitly approved, pseudo text, watermarks.
Output: 9:16, high fidelity, immediately understandable without voiceover.
```

## Quality Checklist

- The first frame shows a concrete viewer problem without voiceover.
- The red flag is visible or can be rendered deterministically on-screen.
- The scene is not generic laptop B-roll or abstract glassmorphism.
- The visual would still make sense to a non-technical viewer.
- It is safe to send to parents/partner/colleagues.
- It leaves safe lower space for active-word subtitles.
- It contains no real brands/logos/QR/bank data/phone numbers/domains.

## Rejection Criteria

Reject if:

- Pretty Empty Visual Blocker: premium image, no visible problem.
- It needs a voiceover to explain what is at stake.
- It contains fake-real text, pseudo-writing, QR codes, brand marks, or financial/legal details.
- It looks like an operational scam template.
- It cannot support a simple “one safer move”.
- It is not mobile-legible at 9:16.

## Provider Policy

- `allowed_image_provider = openai-codex`
- `preferred_image_model = gpt-image-2-high`
- `preferred_quality = high`
- max 3 styleframes per scene without explicit approval
- max 1 retry per styleframe
- no random batches

## Text Rule

For this strategy, screen text is usually deterministic. If the user explicitly approves AI-generated readable screen text, it must pass visual QA and must not contain real brands, real phone numbers, real domains, QR codes, bank data, or copied scam examples.