banana-claude

---
name: banana-claude
description: Use when generating or editing images with Gemini Nano Banana models — acts as a Creative Director that interprets intent and builds an optimized prompt via a 5-component formula, never passing raw text to the API.
source: https://github.com/AgriciDaniel/banana-claude
generated: 2026-07-02T18:41:22.762Z
category: tool
audience: creators
---

## When to use

- Generating an image from an idea where the raw request needs to be interpreted and enhanced before hitting the API
- Editing or doing multi-turn creative sessions with character/style consistency via the Nano Banana MCP
- Selecting a domain lens (cinema, product, portrait, editorial, UI, logo, landscape, infographic) to shape prompt emphasis
- Applying brand/style presets and tracking generation cost across batch variations
- Producing photorealistic ads or commercial product renders that need real camera/brand anchors

## Key concepts

### Creative Director pipeline

Never pass the user's raw text to gemini_generate_image. Always analyze intent, pick a domain mode, construct an optimized prompt, select model + imageSize, then generate — confirming with the user when the request is ambiguous.

### 5-Component Formula

Build every prompt as Subject → Action → Location/Context → Composition → Style (incl. lighting). Be specific and visceral — describe what the camera sees, not what the ad means.

### Domain modes

Choose an expertise lens (Cinema, Product, Portrait, Editorial, UI/Web, Logo, Landscape, Abstract, Infographic); each shifts prompt emphasis — e.g. Cinema → camera/lens/film stock, Product → materials/studio lighting/clean BG.

### Concrete anchors over hype

Name real cameras ('Sony A7R IV'), real brands for styling, micro-details ('sweat droplets on collarbones'), and prestige context ('Vanity Fair editorial'). Never use banned words like '8K' or 'masterpiece' — set imageSize instead.

### Safety + retry loop

On IMAGE_SAFETY apply a safety rephrase and retry (max 3, with user approval); on 429 back off; on 400 FAILED_PRECONDITION inform about billing and stop. Never report success until a valid image file path exists.

## API reference

```
/banana [generate|edit|chat|inspire|batch|preset|cost] <idea or path>
```

Interactive Creative Director commands — generate with full prompt engineering, edit, run consistent multi-turn sessions, browse ideas, or batch N variations.

```
/banana generate a cozy coffee shop hero image for a blog
/banana edit ./hero.png make it golden hour, keep the layout
/banana batch minimalist logo for a fintech app 3
```

```
5-Component photorealistic/ad template
```

Structure a photorealistic or advertising prompt with camera, lighting, and prestige context.

```
[Subject: age + appearance + expression], wearing [outfit with brand/texture], [action verb] in [specific location + time]. [Micro-detail about skin/hair/sweat]. Captured with [camera model], [focal length] lens at [f-stop], [lighting]. [Prestige context: "Vanity Fair editorial"].
```

## Gotchas

- Read references/gemini-models.md and references/prompt-engineering.md before constructing any prompt — not optional, even for simple requests.
- Never pass the user's raw text straight to the API; always interpret and rebuild it with the 5-component formula.
- Banned keywords ('8K', 'masterpiece', 'ultra-realistic', 'high resolution') hurt output — control fidelity with the imageSize param instead.
- Describe the scene, not the concept — write what the camera sees, never 'a dark-themed ad showing…'.
- Requires the @ycse/nanobanana-mcp server and a Gemini API key; generations are billable and may hit IMAGE_SAFETY moderation.
- Ask clarifying questions for vague requests (use case, style, brand) before generating.

---
Generated by SkillMake from https://github.com/AgriciDaniel/banana-claude on 2026-07-02T18:41:22.762Z.
Verify against source before relying on details.

File: ~/.claude/skills/banana-claude/SKILL.md