# Clean AI Image Cut Production QA — no-overlay final-near renders

Use this when producing TrueTraceShorts / AutoShorts videos after the user rejects render overlays and wants the entire visual content carried by high-quality, detailed AI images.

## Accepted production shape

- Visual layer: detailed AI-generated screen/UI images that themselves carry the story and contain the necessary readable scene text.
- Motion layer: subtle Ken Burns/push-in/crossfade between AI images only.
- Caption layer: active-word subtitles are allowed.
- Forbidden renderer additions: progress bars, cards, labels, arrows, circles, focus lines, boxes, connector lines, diagram markers, or explanatory UI overlays.
- Do not add a subtitle box/bar if the user asked for no render overlays; use text-only subtitles with stroke/shadow over a deliberately calm/dark lower safe zone.

## Prompting pattern

For each keyframe prompt, describe the full scene in detail:

- vertical 9:16, realistic laptop/phone/tablet or digital screen environment
- exact readable UI text that belongs inside the generated image
- surrounding UI details that explain the mechanism naturally: sidebars, tabs, cards, prompts, status chips, document/calendar panels
- first/gallery keyframe: include the scam name inside the depicted screen in consistent readable red text; this is for YouTube gallery/swipe recognition and should normally appear only in the first visible image, not every keyframe
- no people, hands, faces, logos, watermarks, QR/barcodes, or real brand marks unless explicitly required
- lower 25% calm/dark for subtitles
- explicitly say there must be no arrows, circles, focus boxes, or overlay-style annotations

## Rendering pattern

1. Generate 4–8 AI keyframes with very detailed prompts.
2. Build a contact sheet and visually inspect it before/while rendering.
3. Assemble as full-bleed 1080x1920 MP4:
   - cover-crop keyframes to 1080x1920
   - subtle Ken Burns only
   - short image-to-image crossfades at scene boundaries
   - text-only active-word subtitles with shadow/stroke
   - no progress rail or decorative caption container
4. Encode with square pixels and explicit vertical display metadata:
   - `setsar=1,setdar=9/16,format=yuv420p`
   - verify `width=1080`, `height=1920`, `sample_aspect_ratio=1:1`, `display_aspect_ratio=9:16`
5. Extract QA frames from early/middle/end and inspect real MP4 frames, not only the contact sheet.

## Caption timing pitfall

EdgeTTS `WordBoundary` events can strip punctuation, so naive grouping by `word.endswith('.')` may create bad cross-sentence chunks such as `SPOTS THIRD GIVE`.

Fix: align word-boundary tokens back to the source script token positions and use source-token punctuation plus a small phrase-breaker set to end subtitle chunks. Keep chunks short, usually 2–3 words for this user's uppercase active-word style.

## QA checklist

Before delivery, verify and report:

- no non-subtitle render overlays are present
- subtitles are readable, within mobile safe zone, and not awkwardly crossing sentence boundaries
- keyframes look premium and screen/UI text is useful enough to understand the mechanism
- no logos/watermarks/people/hands/faces slipped into the AI images
- final MP4 probes as 1080x1920, SAR 1:1, DAR 9:16
- provide media path, duration, bitrate/size if useful, and SHA256 for review packages
- silent-viewing readability: captions must form meaningful sentence/clause blocks, stay on screen long enough to read, and avoid orphan-word pages; if needed, lengthen the video and reduce caption page churn rather than compressing text into the old duration
- For the approved readable-caption pattern and implementation details, see `references/readable-silent-short-captions.md`.