# TrueTraceShorts format-aware premium-AI render discipline

Use this when the user asks for the next TrueTraceShorts / AutoShortsBot video after the format-family expansion.

## Hard correction captured

The user explicitly rejected falling back to renderer-built phone/chat/app graphics after the clean premium-AI workflow had been established. A technically valid render is not review-ready if the main visuals look like self-made PowerPoint/mockup screens.

## Positive reference baseline

Treat `tt003-true-scam-story-fake-support-popup` as a positive reference example for the new quality standard after TT002 correction: premium AI keyframes, no PowerPoint/mockup final graphics, clear `TRUE_SCAM_STORY` family, real everyday situation, safe `generic_pattern` / `educational_pattern` claim mode, no invented victim details, no real numbers/domains/logos/bank data, Gianna Premium Clone v1, clean static active-word captions, and YouTube/TikTok copy. Do **not** treat TT003 as an error case.

## Simplified user command mode

See also:

- `references/truetrace-ok-produce-marketplace-quiz-lessons.md` for the validated OK_PRODUCE production flow, QA gates, dry-run env var, website schema pitfall, and final report shape.
- `references/truetrace-marketplace-quiz-audio-upload-lessons.md` for marketplace quiz CTA wording, Gianna audio loudness/metallic-artifact prevention, caption sentence-boundary repair, and upload-only/publication policy.

Default interaction mode uses three commands:

- `GO_NEXT_SHORT`: propose exactly one next main topic and at most two short alternatives. Auto-check recent produced videos, format mix, topic registry, existing scripts/assets, website companion target, and variety. Reply only in this compact shape: Format, Thema, Hook, Warum jetzt, Website-Ziel, Länge, Erwarteter Nutzen, Alternativen 1/2. No images, no render, no long review package.
- `NEXT_TOPIC`: propose a truly different topic, not a slight variation: different scam type, different format family, different user context.
- `OK_PRODUCE`: produce the currently proposed topic end-to-end, while still enforcing all gates: existing asset check, script/source selection, premium AI keyframes, keyframe QA, Gianna Premium Clone v1, approved captions, render, upload dry-run, title/description, TikTok caption, website companion/link. Stop and report blockers if a gate fails.

After TT003, prefer a non-story/non-support next proposal unless user overrides: Already Clicked fake delivery link, Marketplace off-platform quiz, Overpayment breakdown, or Hi Mom/new-number red flag.

## Required sequence before creating the next video

1. **Load the format-family context first**
   - Red Flag Short
   - Already Clicked?
   - Spot the Trap Quiz
   - True Scam Story
   - Scam Breakdown

2. **Choose the format before choosing visuals**
   - Do not default to another Red Flag or Delivery theme just because a renderer exists.
   - Maintain the intended mix: roughly 50% Red Flag, 20% Already Clicked, 15% Story, 10% Quiz, 5% Breakdown.
   - If the last produced videos clustered around one theme/mechanic, deliberately vary the next one unless the user requested continuity.

3. **Audit existing scripts/artifacts before writing a new renderer**
   - Search for existing render scripts and post-candidates for the topic.
   - If a script already exists with `AI_SCENE_SOURCES` / AI-keyframe workflow, adapt that instead of building a new deterministic UI renderer.
   - Preserve known-good caption/forced-alignment/package logic from previous renderers.

4. **Premium visual rule**
   - Final main visuals must be high-quality AI keyframes with explicit scene descriptions: what is visible, viewer context, scam moment, no real brand/domain/phone/QR/bank/card data.
   - Renderer may crop, apply subtle motion, assemble, and add text-only active-word subtitles.
   - Renderer must not create final phone/chat/app/browser screens from boxes, labels, flat cards, or PowerPoint-like UI.
   - Deterministic UI can exist only as a hidden blueprint, controlled tiny element, or internal prototype — not as the final main image layer.

5. **First-frame rule remains active**
   - Scam name appears only in the first visible image, in red, inside the screen/image itself.
   - Later scenes should not add red labels, circles, arrows, or explainer overlays unless explicitly approved.

6. **Story/claim discipline**
   - For True Scam Story, do not invent concrete victims, amounts, dates, or real case details.
   - Use `common scam pattern`, `this is how it often unfolds`, or equivalent educational-pattern wording.

7. **QA gates before delivery**
   - Vision-QA the AI keyframes before rendering.
   - Vision-QA contact sheet and extracted real frames after rendering.
   - Treat duration windows as soft targets. A few extra seconds are acceptable when they improve sentence-end emphasis, comprehension, or CTA clarity; do not rush voiceover just to hit an exact target.
   - For CTAs, prefer a concrete answerable question (e.g. whether this scam was tried on the viewer or someone they know) over vague prompts like "comment what you noticed first".
   - Audio QA is mandatory for Gianna renders: avoid `speed_factor` stretching because it can create metallic/blechernen tone; measure loudness after render; normalize voice to a Shorts-safe level (around -18 LUFS integrated, true peak about -1.5 dBFS) unless a specific platform target overrides it.
   - Block and rerender for: PowerPoint/mockup look, AI pseudo-brands/domains/phone/QR/bank/card data, caption overlap, ghosting/crossfade double text, clipped text, wrong duration window, stale approval hashes.
   - If a render is rejected for visual standard, mark it not release-ready/revert or quarantine its renderer so the bad workflow does not become reusable precedent.

## Good repair pattern from the session

Bad path: created a Spot-the-Trap Delivery render with renderer-built UI screens. User rejected it because it violated the premium AI-keyframe standard and repeated an old theme.

Correct path:
- revert/quarantine the bad renderer;
- update skill references to make the visual rule a hard blocker;
- choose a different format family (`True Scam Story`) and topic (`fake support popup`);
- generate premium AI keyframes with explicit visible-scene prompts;
- adapt an existing AI-keyframe renderer rather than making new UI screens;
- use Gianna Premium Clone v1, forced alignment, two-line active-word subtitles;
- dry-run upload package and full tests;
- deliver MP4 + thumbnail + title/description + TikTok caption + exact hash-bound approval command.