image-gen

AI image generation for coding agents. Craft optimal prompts, generate images via Google's Nano Banana (Gemini) models, self-review output, and post-process — all from your AI coding assistant.

Works with Claude Code, Codex CLI, Gemini CLI, GitHub Copilot, Cursor, OpenCode, and any tool supporting the SKILL.md open standard.

Quick Start

Install

Pick your platform:

Claude Code (recommended):

# 1. Add the marketplace
claude plugin marketplace add harshkedia177/image-gen-plugin

# 2. Install the plugin
claude plugin install image-gen

Cross-platform (Codex, Gemini CLI, Copilot, Cursor, and more):

npx skills add harshkedia177/image-gen-plugin

Claude Code (local dev / one-off):

git clone https://github.com/harshkedia177/image-gen-plugin.git
claude --plugin-dir ./image-gen-plugin

Manual install (any agent):

# Copy the cross-platform skills into your project
git clone https://github.com/harshkedia177/image-gen-plugin.git
cp -r image-gen-plugin/.agents/skills/* .agents/skills/
rm -rf image-gen-plugin

Set Up Your API Key

Get a free API key from Google AI Studio and set it:

export GEMINI_API_KEY="your-api-key-here"

Add it to your shell profile (~/.zshrc, ~/.bashrc) to persist across sessions.

Generate Your First Image

Just ask your coding agent:

Generate an image of a golden retriever puppy sitting in a sunlit meadow

Or use the slash command in Claude Code:

/image-gen:generate a cyberpunk city at night with neon reflections

That's it. The agent handles prompt crafting, model selection, API calls, and saving the output.

---

What It Does

When you ask for an image, the skill:

1. Gathers requirements — asks 2-3 smart clarifying questions (style, mood, composition) or skips straight to generation if your description is detailed enough
2. Selects the prompting strategy — natural language for simple requests, JSON structured prompts for complex multi-element scenes
3. Picks the right model — Nano Banana 2 for speed, Pro for maximum quality
4. Crafts an optimized prompt — applies Gemini-specific best practices (priority ordering, constraint handling, text rendering tricks)
5. Generates the image — calls the Gemini API with proper imageConfig (resolution, aspect ratio, thinking mode)
6. Self-reviews — evaluates the output for prompt adherence, technical quality, and composition; regenerates with targeted fixes if needed
7. Post-processes — removes backgrounds, converts formats if requested
8. Saves with metadata — image file + markdown sidecar with the exact prompt, settings, and review notes

Features

Feature	Description
Smart prompt crafting	Natural language or JSON structured prompts based on complexity
Two models	Nano Banana 2 (fast, ~3s) and Pro (max quality, ~10-20s)
Thinking mode	Configurable reasoning for complex scenes (NB2 only, 3 levels)
Search grounding	Google Search for real-world accuracy — landmarks, logos, products (NB2 only)
Self-review loop	Automatically evaluates and improves output based on complexity
Conversational editing	Refine images with follow-ups instead of regenerating from scratch
Batch generation	Multiple variations from one concept — vary style, lighting, aspect ratio
Text rendering	Text-first workflow for accurate in-image text (signs, posters, logos)
Character consistency	360-degree sheets, identity locking for multi-image character work
Background removal	Transparent PNGs via rembg — great for logos and stickers
Format conversion	Export as PNG, JPEG, or WebP
14 aspect ratios	From 1:1 to 21:9 ultrawide, 9:16 stories, and more
5 resolutions	512px preview to 4K print-ready
Markdown sidecars	Every image saved with its prompt, model, settings, and review notes

Models

Model	API ID	Speed	Cost	Best For
Nano Banana 2	gemini-3.1-flash-image-preview	~3s at 1K	~$0.067/2K	Fast iteration, batch, web-grounded generation
Nano Banana Pro	gemini-3-pro-image-preview	~10-20s at 1K	~$0.134/2K	Maximum quality, complex scenes, professional output

Fallback: gemini-2.5-flash-image if primary models return 404. Auto-discovers latest models as a last resort.

Usage Examples

Simple generation:

Generate a watercolor painting of a lighthouse at sunset

With specific requirements:

Create a 16:9 product photo of a matte black coffee mug on a marble countertop,
soft morning light from the left, shallow depth of field, no text

Batch generation:

/image-gen:batch Generate 4 variations of a minimalist logo for "Brew Lab" coffee shop
— vary the style: line art, geometric, hand-drawn, modern sans-serif

With post-processing:

Generate a cartoon mascot character on a white background, then remove the background
to make it a transparent PNG

Conversational editing:

> [generates initial image]
Make the lighting warmer and add a slight lens flare from the top right

Prompt optimization (fix a bad prompt):

Optimize this image prompt: "cat, cute, fluffy, orange, garden, sunny, sitting, flowers"

Convert from another model:

Optimize this Midjourney prompt for Nano Banana: "cyberpunk city --ar 16:9 --v 6 --style raw"

Prerequisites

• GEMINI_API_KEY — required. Get one free at Google AI Studio
• Python 3.9+ with rembg — only needed for background removal:

  pip install rembg

Settings (Claude Code)

Create .claude/image-gen.local.md in your project to customize defaults:

---
output_dir: ./generated-images
default_model: nano-banana-2
default_resolution: 1K
default_aspect_ratio: 1:1
review_rounds: 0
save_prompt: true
thinking_level: auto
search_grounding: false
---

Setting	Options	Default	Description
output_dir	Any path	./generated-images	Where images are saved
default_model	nano-banana-2, nano-banana-pro	nano-banana-2	Generation model
default_resolution	512px, 0.5K, 1K, 2K, 4K	1K	Output resolution
default_aspect_ratio	14 ratios (see below)	1:1	Output aspect ratio
review_rounds	0-3	0 (auto)	Self-review iterations. 0 = auto based on complexity
save_prompt	true, false	true	Save markdown sidecar alongside image
thinking_level	auto, off, moderate, advanced	auto	NB2 thinking mode. Advanced costs 20-40% more
search_grounding	true, false	false	Enable Google Search for real-world subjects (NB2 only)

Aspect ratios: 1:1 2:3 3:2 3:4 4:3 4:5 5:4 9:16 16:9 21:9 1:4 4:1 1:8 8:1

Commands (Claude Code)

Command	Description
/image-gen:generate [description]	Generate a single image. Interactive if no description provided.
/image-gen:batch [concept]	Generate multiple variations from one concept.

Skills Included

Skill	What It Provides
image-generation	Complete end-to-end generation workflow (universal, works on all platforms)
prompt-optimizer	Takes rough/bad prompts and rewrites them — diagnoses issues, applies Gemini best practices, recommends settings
nano-banana-prompting	Prompt engineering expertise — NL vs JSON strategy, best practices, examples
image-post-processing	Background removal and format conversion

Each skill includes detailed reference files for advanced techniques, character consistency, JSON schema, prompt examples, and full API documentation.

Architecture

image-gen-plugin/
├── .agents/skills/           # Cross-platform skills (Codex, Gemini CLI, Copilot)
│   ├── image-generation/     # Universal end-to-end workflow
│   ├── prompt-optimizer/     # Prompt diagnosis and rewriting
│   ├── nano-banana-prompting/# Prompt crafting knowledge + references
│   └── image-post-processing/# Background removal & format conversion
├── .claude-plugin/           # Claude Code plugin manifest
├── commands/                 # Claude Code slash commands
│   ├── generate.md           # /image-gen:generate
│   └── batch.md              # /image-gen:batch
├── agents/                   # Claude Code subagents
│   ├── prompt-crafter.md     # Prompt engineering specialist
│   └── image-reviewer.md     # Quality review specialist
├── skills/                   # Claude Code plugin skills
│   ├── prompt-optimizer/
│   ├── nano-banana-prompting/
│   └── image-post-processing/
└── scripts/
    └── remove-bg.py          # Background removal script (rembg)

How It Works Under the Hood

The plugin calls Google's Gemini API via curl with structured imageConfig parameters for reliable resolution and aspect ratio control. It uses a 3-step model resolution chain:

1. Primary model (gemini-3.1-flash-image-preview or gemini-3-pro-image-preview)
2. Fallback (gemini-2.5-flash-image) if primary returns 404
3. Auto-discover latest available image models as last resort

Prompts are crafted following Gemini-specific best practices:

• Priority-ordered elements (subject first, constraints last)
• JSON structured prompts for complex scenes to prevent concept bleeding
• Text wrapped in quotes for in-image rendering
• imageConfig for resolution/aspect ratio (not prompt text)
• Thinking mode for spatial reasoning in complex compositions
• Search grounding for real-world visual accuracy

License

MIT