Google Flow Prompting — Advanced Reference

This skill contains only things the models genuinely require that are non-obvious and specific to these models. No generic advice about "be descriptive."

Nano Banana 2 — Model Architecture You Must Understand

Nano Banana 2 (Gemini 3.1 Flash Image) is not a pure diffusion model. It uses a two-stage pipeline:

1. Reasoning stage — Gemini 3.1's language understanding interprets the prompt first, resolving spatial logic, relationships, and world knowledge before any pixels are generated.
2. Diffusion stage — a dedicated diffusion model converts the reasoning output into the image.