# High-quality AI-image repair after mockup-style rejection

Use this when the user rejects a TrueTraceShorts / AutoShorts render as not good enough and explicitly asks to use the previous/high-quality AI image style again, e.g. `Nutze wieder die hochwertigen KI Bilder!`.

## Trigger

- A render used deterministic mockups, layout cards, or renderer-built screens as the main visual layer.
- The user says the video is not good, too mockup-like, too PowerPoint/wireframe, or asks for high-quality AI images again.

## Required response pattern

1. Treat the comment as a hard visual-style correction, not a minor preference.
2. Do not defend the deterministic mockup approach and do not ask for another topic by default.
3. Reuse the same approved script/topic unless the user asks to change content.
4. Generate or select 4–8 premium AI image keyframes as the main visual layer:
   - vertical 9:16;
   - realistic laptop/phone/screen/work-scene environment;
   - story carried by high-quality AI images;
   - for TrueTraceShorts scam videos, the **first/gallery keyframe only** should include the scam name inside the depicted screen in consistent readable red text, e.g. `REMOTE JOB SCAM`, so the YouTube gallery/swipe overview stays uniform; do not repeat the scam-name label in every later keyframe unless the user explicitly asks;
   - no people/hands/faces/logos/watermarks unless explicitly desired;
   - no renderer-like arrows/circles/boxes/labels/progress bars;
   - lower safe area calm/dark enough for subtitles.
5. Keep the renderer limited to subtle Ken Burns/crossfades and text-only active-word subtitles.
6. Re-render the MP4 and rebuild all review/upload artifacts because the video hash changed.
7. Re-run visual QA on contact sheet and extracted subtitle frames before delivery.
8. Deliver a fresh review package with the new video SHA256, posting-pack SHA256, and updated hash-bound approval command.
9. If the user approves in natural language immediately after that one package, reuse the package's stored approval command, run dry-run validation, validate upload-only scope, then execute private upload.

## Pitfalls

- Do not leave the old deterministic-mockup MP4 in the media cache under the same path without overwriting or clearly replacing it; Telegram delivery must point to the repaired video.
- Do not reuse the previous approval command after visual repair; the video hash must change, so the approval command must change.
- Do not call the repaired style “AI background plus deterministic mockups” if the user specifically rejected mockups. In that case the AI keyframes are the primary visual style; deterministic elements should be absent or minimal.
- If AI-generated screen text is central to the story, inspect it visually. Reject obvious fake text, real logos, watermarks, or clutter that competes with captions.
- For fake CAPTCHA / browser-permission shorts, do not use self-built CAPTCHA graphics as the main final image just to avoid reCAPTCHA/Google logos. Instead, regenerate well-prompted AI keyframes with explicit negatives (`no reCAPTCHA logo`, `no Google branding`, `plain checkbox only`, `generic browser UI`, `no real domains`) and verify the contact sheet. Deterministic UI is acceptable only as a blueprint or tiny controlled layer, not as a PowerPoint-looking final visual.