Prompt Basics
Core concepts for writing prompts that models parse reliably—subject, style, camera, lighting, and constraints.
Principle from Google’s Nano Banana / Gemini image docs
Google’s image generation guide states that the model’s strength is deep language understanding. The guideline is:
Describe the scene, not just a list of keywords. A short narrative paragraph almost always produces better, more coherent images than a string of unrelated words.
That matches how Nano Banana Pro 2 behaves in this app: treat the prompt as a creative brief, not a bag of tags. Official reference: Gemini API — Nano Banana image generation.
The prompt is a creative brief
Think like a brief to a cinematographer: what is in frame, how it is lit, which lens personality you want, and what must never appear.
The five building blocks
| Block | Examples |
|---|---|
| Subject | solo hiker, red jacket, mountain ridge at dawn |
| Scene | fog rolling through valley, distant peaks |
| Camera | 35mm documentary feel, handheld micro-shake |
| Light | backlit rim, cool shadows, warm highlights |
| Finish | subtle film grain, natural color, high clarity |
Order matters less than density
Models read the whole prompt. Put the most important constraints first and repeat critical constraints once later if they are easy to miss (for example exact text on a sign).
Start simple, then specialize
- Get a believable base image with subject + light + camera.
- Add material words (metal, brushed, velvet, matte paint).
- Add composition words (rule of thirds, centered subject, negative space).
Official “best practices” (condensed)
From the same documentation, patterns that reliably improve results:
- Be specific — e.g. ornate elven plate armor with silver leaf etching reads better than “fantasy armor.”
- Give context and intent — e.g. “logo for a premium minimalist skincare brand” beats “make a logo.”
- Iterate in small steps — e.g. “keep everything, but warmer light” or “same scene, stricter expression.”
- Use step-by-step instructions for busy scenes (background first, then foreground, then hero object).
- Semantic negatives — describe the scene you want (e.g. “empty deserted street with no traffic”) instead of only “no cars.”
- Control the shot with photo/cinema terms: wide-angle, macro, low-angle perspective, etc.
Related reading
- Prompt structure for templates and sections (including Gemini-style templates).
- Vocabulary choice for word swaps that change the look.