Image Prompt Skills for Claude Code

One line in, a gold-quality image prompt out. image-prompt turns a short idea into a structured, copy-paste-ready prompt for any major image model: posters, landing and UI mockups, ads, editorial, game art, logos, infographics. You supply the idea; the skill builds the whole prompt. Press c to copy.

No prompt-engineering required.

Highlights

• Nine-block layered prompts — named composition zones, hex palette, real font references, mood, finish, and IP-safety. The pattern behind gold-tier prompts.
• Nine archetypes — poster, landing hero, product ad, UI mockup, photoreal scene, game screenshot, infographic, logo, illustration.
• Eight model adapters — the same idea renders correctly per model, each distilled from the model's best available prompting docs.
• Style menu — say "help me with the style" and pick from a built-in repertoire.
• Private style presets — save a look you love and reuse it on any idea; share it as a portable imgpreset:v1:… code.
• Edit / remix mode — identity-preserving posters, style transfer, background swap, product-on-white, text localization.
• Optional generation — opt-in: render the PNG through a provider API.

Install

Claude Code, plugin marketplace (recommended)

/plugin marketplace add veryCoolTimo/imagegen-skills
/plugin install imagegen-skills@imagegen

Claude Code, manual (clone + symlink)

git clone [email protected]:veryCoolTimo/imagegen-skills.git
cd imagegen-skills
ln -s "$(pwd)/skills/image-prompt" ~/.claude/skills/image-prompt

Restart Claude Code so it discovers the skill. It also works on claude.ai and Claude Desktop: import the repository from the Skills panel. The skill produces text only, so no keys or dependencies are needed (generation is a separate opt-in, see below).

Usage

Describe what you want, in any language:

сделай крутой промт для постера панк-выставки в Берлине, лимитированный тираж

landing hero for a specialty coffee delivery app, warm and minimal

game screenshot of a viking longship raid at dawn, AAA fidelity, for Midjourney

инфографика про наш продукт, в стиле антропик, через nano banana pro

I want a poster — help me with the style

put me on a movie one-sheet as an aquatic superhero (I'll attach my photo)

You get the finished prompt in a copyable block plus a one-line assumptions note (archetype, size, quality, brand, model). Redirect with a single sentence: "make it 9:16", "darker palette", "another variant".

Models

Default target is gpt-image-2. The core builder is model-agnostic: each model is one adapter file under references/models/, so the same idea renders correctly per target. Just name a model and the skill switches adapters.

Style presets and sharing

Found a look you love? Save it as a private preset and reuse it on any future idea.

save this style as neon-noir

A preset captures the style layer only (background, palette, fonts, motifs, mood, finish); content still comes from each new idea. Presets live per-user at ~/.claude/image-prompt/presets/<name>.md, private to your machine and never committed. Reuse:

a poster for a coffee shop, in the neon-noir style

Share a preset with anyone, no hosting or accounts: export preset neon-noir returns a self-contained imgpreset:v1:… code (gzip + base64). Whoever runs import preset <code> gets the exact preset. Manage with "list my presets", "update", "delete".

Optional: generate the image

By default the skill only writes the prompt. If you ask it to generate and a provider API key is set, it renders the PNG via scripts/generate.py and saves it into the project. Supported providers (key in an env var, never committed):

openai · openrouter (universal — one key, many models) · gemini (Nano Banana Pro) · ideogram · recraft · bfl (FLUX.2) · reve. Midjourney has no API, so it stays prompt-only.

python3 scripts/generate.py --provider openrouter --prompt-file prompt.txt \
    --model google/gemini-3-pro-image --aspect 16:9 --resolution 2K --out out.png

Generation is opt-in and paid; the skill confirms before spending.

How it works

idea → detect archetype → expand to the nine blocks → format for the target model → output

The nine blocks: concept line · canvas and background · named composition zones · subject detail · typography (literal copy + real fonts) · hex palette · font role map · mood cluster · finish and quality tag. Photoreal and game-screenshot ideas use a labeled-block variant (scene, subject, lighting, camera, HUD, technical). Each model adapter changes only the rendering, not the content.

Project structure

imagegen-skills/
├── README.md
├── LICENSE                          # MIT
├── .claude-plugin/                  # plugin.json + marketplace.json (powers the marketplace install)
├── assets/                          # banner + the prompt used to make it
├── scripts/generate.py             # optional: render a prompt via a provider API (opt-in)
└── skills/image-prompt/
    ├── SKILL.md                    # trigger, workflow, presets, style menu, edit, output
    └── references/
        ├── anatomy.md              # the nine-block skeleton
        ├── archetypes.md           # templates + default size/quality per archetype
        ├── fonts-palettes.md       # font library by vibe + hex palette discipline
        ├── styles.md               # built-in style menu (repertoire to pick from)
        ├── edit-remix.md           # edit/remix workflow templates
        ├── generation.md           # how the optional generate step calls provider APIs
        ├── gold-examples.md        # curated reference prompts, grows over time
        └── models/                 # one adapter per model (distilled from each model's best docs)
            ├── gpt-image-2.md      # + gpt-image-1 / 1.5 family
            ├── gemini.md           # Nano Banana Pro (Gemini 3 Pro Image)
            ├── midjourney.md
            ├── flux.md
            ├── recraft.md
            ├── ideogram.md
            ├── reve.md
            └── universal.md

Private, per-user data (style presets, your UI reference library) lives under ~/.claude/image-prompt/ and is never committed.

Extending

• New favourite prompt: add it to references/gold-examples.md with a Teaches: tag.
• New model: add references/models/<name>.md; the workflow stays unchanged.
• New style: add it to references/styles.md, or save a private preset and share the code.

License

MIT