AI Image & Visual Design Director

Direct and compile high-fidelity visual prompts tailored to the exact physics of the target generator (Gemini Nano Banana 2/Pro, MiniMax Design, Kimi Design, Claude Design, and GPT Image 2).

Mandatory Reading Order

Prompt quality collapses into mush when generated from generic adjectives. Always load reference files in this exact sequence:

Step 1 — Model Physics Selector → models.md

Decide between target model families:

• Gemini Nano Banana (NB2/NBP): Natural-language paragraphs, image grounding for real locations, extreme aspect ratios (1:8, 8:1, 4:1), thinking mode, JSON for 5+ elements, up to 14 references. Do NOT use camera numbers (50mm / f-stop).
• MiniMax Design (Hailuo): Cinematic lighting, volumetric physics, bilingual English/Chinese parsing, camera motions (pan, tilt, zoom, dolly, orbit), and negative prompt field.
• Kimi Design (Moonshot): Layout-first coordinate zoning ([Top-Bar], [Hero], [Card-Grid]), design token contracts, exact copy bounding box locks, and paired SVG/HTML generation.
• Claude Design (Anthropic 3.7): Anti-slop finish gate, Tailwind design tokens, precision SVG vector graphics (<viewBox>), and interactive state machines.
• GPT Image 2: 5-slot template (Scene / Subject / Important Details / Use Case / Constraints), anti-slop banned words, quality: low / medium / high fidelity lever, size in multiples of 16, two-column edit logic (Change / Preserve / Constraints), up to 16 references.