# Everyday Red Flags — clean AI image render recipe

Use this when producing the user's current AutoShorts/TrueTraceShorts format: realistic everyday scam videos for normal people, built from high-quality AI images plus text-only active-word subtitles.

## Default production shape

1. Pick a broad, concrete scam that normal people recognize:
   - delivery fee SMS
   - “Hi Mom / new number” emergency text
   - fake bank/login message
   - fake support pop-up
   - marketplace off-platform payment
   - changed invoice details
2. Write a short plan JSON or equivalent with:
   - `candidate_id`
   - hook
   - red flag
   - safe action
   - takeaway
   - final English voiceover script
   - 6-scene image plan
3. Generate exactly enough realistic 9:16 image keyframes for the story, typically 6:
   - scene 1: scam visible instantly
   - scene 2: why it feels believable
   - scene 3: red-flag combination
   - scene 4: safe action 1
   - scene 5: safe action 2 / verification
   - scene 6: takeaway / share cue
4. Render with:
   - EdgeTTS English voiceover, usually `en-GB-RyanNeural`
   - EdgeTTS `WordBoundary` events for active-word subtitles
   - text-only subtitles: white words, current spoken word yellow, black stroke only
   - subtle Ken Burns/crossfade motion only
   - no caption box/bar, no progress bar, no visual labels
5. Produce review package:
   - final MP4 path
   - contact sheet
   - extracted QA frames
   - ffprobe metadata
   - SHA256
   - German approval notes, English final copy

## Image prompt rules

Prompts should be very detailed and concrete. They must describe the everyday physical scene and the exact readable in-image phone/screen text that carries the scam.

Include in every prompt:
- `Vertical 9:16 realistic premium AI image`
- ordinary home/work context, not abstract cyber visuals
- no people, no hands, no faces
- no logos, brands, watermarks, QR codes, real domains, or real private data
- exact readable UI/SMS/chat text using generic sender names and example/reserved data
- lower 25% calm/dark for subtitles
- explicit ban on arrows/circles/boxes/warning labels/overlay graphics

Good first-frame examples:
- phone SMS from `Delivery Notice`: `Your package is held. Pay $1.99 delivery fee today.`
- phone SMS from `New Number`: `Hi Mom, it’s me. I lost my phone.` / `Can you send money today? It’s urgent.`

## Caption segmentation pitfall

EdgeTTS `WordBoundary` events may strip punctuation, so do not segment only by the emitted word tokens. Reattach punctuation by aligning word events back to the source script tokens or add phrase breakers. Otherwise captions can cross sentence boundaries and look ugly.

Use short 1–3 word chunks for this style. Avoid long 3–7 word chunks if they crowd the bottom safe zone.

## QA gate

Before delivery, always verify:

- contact sheet: first image instantly shows the scam
- contact sheet: all scenes are realistic everyday scenes, not abstract tech art
- contact sheet: no people/hands/faces/logos/watermarks
- rendered frames: no non-subtitle overlays, boxes, bars, arrows, circles, or progress indicators
- subtitles: readable, not clipped, inside mobile safe zone
- ffprobe: `1080x1920`, `sample_aspect_ratio=1:1`, `display_aspect_ratio=9:16`, 30fps preferred
- review package: German approval summary, English final audience copy, manual approval only

## Blocking rule

If the first frame does not make the scam obvious without hearing the voiceover, regenerate the first image before rendering. Pretty but vague is a failure for this series.