Skip to content

Latest commit

 

History

History
217 lines (155 loc) · 9.9 KB

File metadata and controls

217 lines (155 loc) · 9.9 KB

Nano Banana Prompt Generator - Claude Skill

A Claude skill for generating professional Nano Banana / Gemini image prompts. Works with Claude.ai, Claude Desktop, and Claude Code. Supports Image Generation, Photo Editing, YouTube Thumbnails, Infographics, Storyboards.

Live Generator Claude Skill License: MIT

What is this?

A skill for Claude that helps you write effective prompts for Nano Banana - Google's family of AI image generation models in the Gemini API. Unlike most AI image tools that expect tag lists, Nano Banana works like a creative director: it understands intent, context, and full-sentence descriptions.

Works with Claude.ai (web), Claude Desktop, and Claude Code - all three support the same skill format.

The skill covers all four current Gemini image models:

Marketing name Model code Status Niche
Nano Banana 2 gemini-3.1-flash-image Stable (GA 2026-05-28) Default for everything; only model with video input
Nano Banana 2 Lite gemini-3.1-flash-lite-image Stable (GA 2026-06-30) Sub-2s latency, cheapest per image, 1K only
Nano Banana Pro gemini-3-pro-image Stable (GA 2026-05-28) Studio-quality precision, dense accurate text
Nano Banana (legacy) gemini-2.5-flash-image ⚠️ Shutdown 2026-10-02 Migrate off it

Heads-up if you installed v1.1.0: the -preview model IDs it recommended were shut down by Google on 2026-06-25, and the model it set as the production default is being shut down on 2026-10-02. v1.2.0 fixes both. Update the skill.

Every model spec, capability, and price in this skill is verified against ai.google.dev/gemini-api/docs. See SKILL.md "Verification & Sources" for the full URL list.

Try the interactive web generator (no install): maciejdzierzek.com/narzedzia/generator-promptow

Features

  • 5 generation modes: Image Generation, Photo Editing, YouTube Thumbnails, Infographics, Storyboards
  • Per-model guidance across all four models, including when Lite is the wrong choice
  • Conversational editing patterns - refine iteratively without re-rolling
  • Multi-character consistency - up to 4 characters (NB2) with documented per-model limits
  • Multi-reference compositions - up to 14 reference images with role assignment
  • In-image text translation - preserve design while changing language
  • Image Search Grounding - pull real, current data into infographics (NB2 and Pro)
  • Video-to-image - generate a thumbnail or poster straight from a YouTube URL (NB2 only)
  • Official language list for in-image text, and the fallback when your language is not on it (Polish is not)
  • Storyboard frames for Veo handoff - prompt patterns for video-pipeline workflows
  • Verified pricing reference - per-tier (Standard / Batch / Flex / Priority) costs
  • Trigger accuracy eval suite for the skill-creator framework

Installation

Claude.ai and Claude Desktop (all plans, including free)

  1. Download nano-banana-prompt-generator-skill.zip from this repository (click → Download raw file)
  2. In Claude.ai or Claude Desktop: go to Customize > Skills → click "+""Upload a skill"
  3. Upload the ZIP file and toggle the skill on

Claude will automatically use the skill when you describe a Nano Banana image task. Available on Free, Pro, Max, Team, and Enterprise plans.

Note: Skills require code execution to be enabled. If you don't see the Skills section, go to Settings > Capabilities and enable "Code execution and file creation" first.

Claude Code (CLI) - via plugin marketplace

/plugin marketplace add maciejdzierzek/claude-plugins
/plugin install nano-banana-prompt-generator@maciejdzierzek

One marketplace, all my skills - see maciejdzierzek/claude-plugins. No download needed.

If you added this repository as a marketplace before August 2026, switch to the command above. Claude Code registers one marketplace per name, so the old per-repo catalogues (which all used the same name) replaced each other when you installed more than one skill. This repo's catalogue still resolves, under the name maciejdzierzek-nano, but it is no longer the recommended path.

Claude Code (CLI) - manual install

mkdir -p ~/.claude/skills/nano-banana
cp -r skills/nano-banana/* ~/.claude/skills/nano-banana/

For project-specific use:

mkdir -p /your-project/.claude/skills/nano-banana
cp -r skills/nano-banana/* /your-project/.claude/skills/nano-banana/

Reload Claude Code. The skill activates automatically.

Interactive web generator (no installation)

Use directly at: maciejdzierzek.com/narzedzia/generator-promptow No API key, no registration, no data sent to servers. Works in your browser.

Usage

Describe what you want - the skill picks the right model and template:

Image Generation:

Generate a Nano Banana prompt for a product photo of my coffee grinder.
Minimalist studio look, white background, space for text on the right.

Photo Editing:

I have a photo of a storefront in summer. Write a Nano Banana prompt
to transform it into winter while keeping the building exactly the same.

YouTube Thumbnail:

I need a thumbnail for my iPhone review video. Excited-person-on-left,
product-on-right layout. Help me write it.

Infographic with current data:

Create a Nano Banana Pro prompt for an infographic showing this quarter's
top-5 economies' GDP growth. Use real, current data via search grounding.

Storyboard for Veo:

Write Nano Banana prompts for 6 storyboard frames showing the same character
walking through a forest as the seasons change. Keyframes for Veo handoff.

Model Selection (verified against per-model spec pages)

Use case Recommended model Why
Anything, unless a reason says otherwise Nano Banana 2 Google's own go-to pick: 0.5K-4K, web + image grounding, 14 ratios, up to 14 references
Thumbnail or poster from a finished video Nano Banana 2 The only model that accepts video input
Bulk runs, interactive edit loops, tightest budget Nano Banana 2 Lite Sub-2s latency, $0.0336 / image at 1K (~$0.017 Batch). 1K only, no grounding
Studio-quality work, dense accurate text, brand precision Nano Banana Pro Highest fidelity, search-grounded, Thinking on by default, ~2x Nano Banana 2

Pricing details: see skills/nano-banana/references/pricing.md.

Prompt Structures

Image Generation

[Subject + adjectives] doing [action] in [setting].
[Composition/camera angle]. [Lighting/atmosphere].
[Style/medium]. [Technical specs: resolution, aspect ratio].
[Constraints or text to render if any]

Photo Editing

[Action verb] + [specific element] + [desired result].
KEEP UNCHANGED: [list of what must stay the same].
[Style/quality requirements].

YouTube Thumbnail

YouTube thumbnail for [TOPIC]:

SUBJECT: [person + expression + position in frame]
PRODUCT/VISUAL: [what to show on the other side]
TEXT: "[HEADLINE]" [font style, size, color, position]
BACKGROUND: [treatment]

TECHNICAL: 1280x720px, 16:9, [contrast/color notes]

Infographic

[Type] infographic: "[TITLE]"

STRUCTURE:
- [Section 1 description]
- [Section 2 description]

DESIGN:
- Colors: Primary [HEX], Secondary [HEX], Background [HEX]
- Typography: [font style], headlines [size]pt

TECHNICAL: [resolution], [aspect ratio], all text ultra-sharp
AVOID: Drop shadows, 3D effects, gradients on charts

Examples

See examples/ for ready-to-use prompts:

Complete template library: skills/nano-banana/references/prompt-examples.md.

Golden Rules

  1. Creative Director, not keyword vomit - Write full sentences, not tag lists like dog, park, 4k, realistic, sunny
  2. Edit, don't re-roll - If an image is 80% right, ask for specific changes instead of regenerating
  3. Explain the logic - The model understands intent. Say why, not just what
  4. Specify preservation explicitly - Always state what must not change during edits
  5. Provide context and purpose - "For a Brazilian gourmet cookbook" helps the model make better artistic decisions
  6. Use positive framing - Describe what you want, not what you don't want
  7. Specify HEX codes for brand colors - "background color: #0066CC" is more reliable than "dark blue"
  8. Match model to task - thumbnails on Pro = wasted money; many references or grounded facts on Lite = wrong tool
  9. Never paste a -preview model ID - all of them are shut down; four live IDs only

About the Author

Built by Maciej Dzierżek - consultant, trainer and creator from Poland. Specializes in helping businesses implement AI in day-to-day workflows.

License

MIT - use freely, attribution appreciated.