# Premium AI Styleframe Direction for Hybrid Motion Shorts

Use this when deterministic layout proofs are useful as mechanism blueprints but rejected as too wireframe/PowerPoint for final social-video style.

## Core lesson

Do not swing between two bad extremes:

- **Bad:** ask AI image/video generation to invent the whole explanatory mechanism; this produces random abstract symbolism.
- **Also bad:** treat deterministic Pillow/layout proofs as the final visual look; this can read as PowerPoint/wireframe.

Preferred pattern:

```text
mechanism blueprint → tightly briefed premium AI styleframes → user selects style → matching start/end keyframes → short Wan/I2V proof → full final only after proof acceptance
```

## Responsibility split

- **Deterministic renderer owns:** captions, labels, arrows, facts, stamps, sorting paths, visual state changes, final assembly, checksums, delivery metadata.
- **AI image generation owns:** premium visual atmosphere, materials, lighting, depth, coherent text-free styleframes/keyframes.
- **Wan/I2V owns:** motion between accepted style/keyframes and subtle background/scene movement.
- **Never let AI own:** factual text, generated labels, fake UI, logos, or the precise explanatory logic.

## Approval gates

Before AI styleframes:

1. Reject/audit any prior layout direction if it is not visually premium enough.
2. Mark deterministic layouts as `composition_blueprint_only`, not final style.
3. Require explicit approval such as:

```text
APPROVED_FOR_AI_STYLEFRAME_DIRECTION <candidate_id> <version> <script_hash>
```

Allowed under that approval:

- maximum 3 AI styleframes by default;
- a 4th only if one direction clearly fails;
- contact sheet;
- Telegram/media delivery with hashes.

Still blocked:

- Wan/I2V;
- T2V;
- endframes;
- full-final render;
- posting/platform APIs;
- render series.

## Prompt shape

Each styleframe prompt should include:

- exact composition;
- style / art direction;
- lighting;
- mobile-first 9:16 camera framing;
- caption-safe area;
- renderer-owned overlay note;
- negative prompt.

Reusable negative prompt:

```text
readable text, letters, numbers, fake UI, logos, brand marks, watermarks, captions, subtitles, document text, charts with labels, people, hands, faces, clutter, messy office, dark sci-fi tunnel, toy miniature look, PowerPoint diagram, low quality, blurry, random abstract shape
```

## Example: meeting summary / decision target mechanism

Brief:

- left/upper-left: 12–18 blank information cards / note fragments;
- center: transparent compression funnel;
- lower-center: smooth generic summary capsule/orb/output container;
- upper-right: empty decision-target node/ring, visible but unconnected;
- bottom 25%: calm caption-safe area.

Variants:

- **A — Premium Glassmorphism Light:** bright Apple/SaaS-style, white/pale-blue, clean glass, very calm.
- **B — Cinematic Work-Tech:** subtle blurred work/meeting context, foreground mechanism sharp, premium product-film feel.
- **C — High-Contrast Social Premium:** dark navy/deep graphite, tasteful glow, stronger Shorts punch, but avoid sci-fi tunnel drift.

## Quality gate

Auto-reject or revise a styleframe if:

- readable text, pseudo-writing, logos, UI, or watermarks dominate;
- cards/funnel/target node are missing;
- decision target is not visibly separate/unconnected;
- mechanism is not roughly understandable without voiceover;
- bottom caption-safe area is missing;
- it looks like PowerPoint/wireframe;
- it is random abstract AI symbolism;
- it is too dark, too busy, toy-like, or stock-office.

## Delivery requirements

For each artifact report:

- `variant_id`;
- `file_path`;
- `file_exists`;
- `file_size`;
- `sha256`;
- `media_type`;
- `telegram_delivery_attempted`;
- `telegram_delivery_verified`;
- `telegram_message_id` if available;
- concise own evaluation and recommendation.

If media is not actually sent/verified, say **local erzeugt, Telegram-Versand nicht erfolgt**. Do not write “attached” unless the platform delivery attempt succeeded.

## Next commands

After delivery, accept only these next-step commands:

```text
APPROVED_PREMIUM_STYLEFRAME_DIRECTION <candidate_id> <A|B|C|D>
REVISE_PREMIUM_STYLEFRAME_DIRECTION <candidate_id> <A|B|C|D> "..."
REJECT_PREMIUM_STYLEFRAME_DIRECTIONS <candidate_id> "..."
```

No video or endframe should start until a premium styleframe direction is approved.

## If the premium pass is still too abstract

If the user says the visual quality improved but the motif still feels too abstract, do **not** generate more empty-background funnel/ring variants. Treat that as a shift in art direction, not a total rejection of AI images.

Next step: use `references/real-life-premium-styleframes.md`.

Core change:

```text
abstract premium SaaS symbol → realistic work-scene base + subtle renderer-owned mechanism overlay
```

Typical allowed directions:

- desk/laptop scene;
- empty meeting room / glass board;
- project table / work evidence.

Keep Wan/I2V/endframes/full-final blocked until the user approves one real-life styleframe direction.