Structured Prompts

# Structured Prompts (JSON / YAML / XML)

**Per-model parsing reliability:**

| Model | JSON | YAML | XML/`<tag>` |
|---|---|---|---|
| GPT Image 2 | ✅✅ best in class — heavy nested objects respected | ✅ | ✅ |
| Nano Banana Pro | ✅ reliable | ✅ reliable | ✅ reliable |
| Nano Banana (standard) | ⚠️ partial | ⚠️ partial | ⚠️ partial |
| GPT Image 1 / DALL·E 3 | ⚠️ partial | ⚠️ partial | ⚠️ partial |
| FLUX [dev/pro] | ⚠️ partial — prefer prose | ⚠️ partial | ⚠️ partial |
| Imagen 3/4 | ⚠️ partial — prefer prose | ⚠️ partial | ⚠️ partial |
| Midjourney v6/v7 | ❌ — flags only | ❌ | ❌ |
| SDXL / SD3.5 | ❌ — tag/weight only | ❌ | ❌ |
| Ideogram 3 | ⚠️ partial | ⚠️ partial | ⚠️ partial |
| Recraft v3 | ⚠️ partial | ⚠️ partial | ⚠️ partial |

Default rule: use structured prompts on Pro-tier natural-language models (GPT Image 2, Nano Banana Pro). Use prose on Standard tier. Use tags/flags on Midjourney/SDXL.

Use structure when:

- The prompt has many orthogonal attributes (subject, wardrobe, lighting, background, etc.)
- You need to reuse the same schema across many generations (templating)
- You want to express conditional rules or nested grouping
- The task is multi-step (analyze → plan → render)

For simple prompts, free-form paragraphs still win.

## JSON Object Format

Best for: character/scene specs with many nested attributes. Seen in 2000s Mirror Selfie, Bathroom Selfie, fashion editorials.