Skills
skills/ericjy/inkframe/prompt-to-image

prompt-to-image

SKILL.md

Prompt To Image

Generate prompt-based images with a bundled Node script that calls provider REST APIs directly.

This skill is for creating new images from prompts, and it also supports basic reference-image-guided generation for specific OpenAI and Gemini models. It is not for complex multi-step compositing workflows or rendering markdown into graphics.

Use This Skill For

  • OpenAI image generation with current GPT Image models such as gpt-image-1.5 (default), gpt-image-1, or gpt-image-1-mini
  • Gemini image generation with current Google image models such as gemini-2.5-flash-image (Nano Banana), gemini-3.1-flash-image-preview (Nano Banana 2), or gemini-3-pro-image-preview (Nano Banana Pro)
  • Reference-image prompting and image edits when the provider/model supports it
  • Single images or small batches of prompt variations
  • Tasks where the user wants a local image file as output

Do Not Use This Skill For

  • Markdown, quote cards, code screenshots, or social post rendering from text blocks Use markdown-to-image instead.
  • Heavy multi-step image-edit workflows, advanced compositing, or provider-agnostic reference pipelines This skill is deliberately simpler than baoyu-image-gen.

Script Path

SKILL_DIR/scripts/main.mjs

Resolve SKILL_DIR as the directory containing this SKILL.md.

Setup

Provide API keys via environment variables:

  • OPENAI_API_KEY for OpenAI models
  • GOOGLE_GENERATIVE_AI_API_KEY for Gemini models

Aliases also supported by the script:

  • GOOGLE_API_KEY
  • GEMINI_API_KEY

Optional default model environment variables:

  • OPENAI_IMAGE_MODEL default: gpt-image-1.5
  • GOOGLE_IMAGE_MODEL or GEMINI_IMAGE_MODEL default: gemini-2.5-flash-image

Optional .env lookup paths:

  • <cwd>/.prompt-to-image/.env
  • ~/.prompt-to-image/.env

The script uses those files as fallbacks beneath existing environment variables, and normalizes the Google key aliases.

No repo dependency changes are required. The script uses Node's built-in fetch.

Quick Start

# Default provider selection prefers Google when both keys are present
node ${SKILL_DIR}/scripts/main.mjs \
  --prompt "A cinematic koi fish temple at sunrise" \
  --output out.png

# Force OpenAI with an explicit image model
node ${SKILL_DIR}/scripts/main.mjs \
  --provider openai \
  --model gpt-image-1.5 \
  --prompt "A paper-cut poster of a lunar rover in a desert" \
  --size 1536x1024 \
  --output rover.png

# Use Nano Banana / Gemini image generation
node ${SKILL_DIR}/scripts/main.mjs \
  --provider google \
  --model gemini-2.5-flash-image \
  --prompt "A charming ramen shop mascot, flat vector style" \
  --ar 4:5 \
  --output mascot.png

# Generate multiple concepts
node ${SKILL_DIR}/scripts/main.mjs \
  --provider openai \
  --prompt "Three different thumbnail concepts for a retro sci-fi podcast" \
  --n 3 \
  --output thumbnails/concept.png

CLI Options

Option Description
-p, --prompt <text> Inline prompt text
--promptfiles <files...> Read prompt text from files and concatenate them
--ref, --reference <files...> One or more reference images
-o, --output <path> Output file path, or basename for multi-image runs
--provider openai|google Force provider selection
-m, --model <id> Model override
--size <WxH> Exact size, for OpenAI image models
--ar, --aspect-ratio <W:H> Aspect ratio. Gemini 2.5 Flash Image supports ratios such as 1:1, 4:5, 9:16, and 16:9 directly
--n <count> Number of images to generate
--quality <value> OpenAI provider option passthrough
--style <value> OpenAI style passthrough, e.g. vivid or natural
--timeout-ms <ms> Abort request after a timeout
--json Print machine-readable metadata after saving files

Provider Selection

The script uses this order:

  1. --provider, if provided
  2. If only one provider key is available, use that provider
  3. If both are available, default to Google because Nano Banana is the more flexible default here

Model ID Cheat Sheet

When choosing a model, prefer the stable, commonly used IDs below unless the user explicitly requests something else.

OpenAI: most common choices

  • gpt-image-1 Use when the user explicitly asks for it or when you need to match the older GPT Image pricing/performance tier.
  • gpt-image-1.5 Default choice for most OpenAI image generation tasks. Use this first for general-purpose text-to-image work.
  • gpt-image-1-mini Use when the user explicitly wants a cheaper/faster GPT Image variant.

Preferred OpenAI default in this skill: gpt-image-1.5.

Google Gemini: most common choices

  • gemini-2.5-flash-image Default Nano Banana model. Use this first for most Gemini image generation tasks.
  • gemini-3.1-flash-image-preview Also known as Nano Banana 2. It provides high-quality image generation and conversational editing at a mainstream price point and low latency. Treat it as the high-efficiency counterpart to Gemini 3 Pro Image when the user wants speed, lower cost, or high-volume developer workflows. Use when the user explicitly asks for Nano Banana 2 or wants a newer Gemini image variant.
  • gemini-3-pro-image-preview Also known as Nano Banana Pro. Use when the user mentions Nano Banana Pro, or wants a more premium or instruction-heavy Gemini image model and explicitly asks for higher-end quality.

Preferred Google default in this skill: gemini-2.5-flash-image.

Selection rules for the LLM using this skill

  1. If the user names a model, use that exact model ID.
  2. If the user asks for OpenAI image generation but does not name a model, start with gpt-image-1.5.
  3. If the user asks for Gemini, Nano Banana, or Google image generation without naming a model, start with gemini-2.5-flash-image.
  4. Only reach for gpt-image-1.5, gpt-image-1-mini, gemini-3.1-flash-image-preview (Nano Banana 2), or gemini-3-pro-image-preview (Nano Banana Pro) when the user asks for them or there is a clear reason to optimize for that tradeoff.

Pricing Notes

Pricing changes over time. If cost matters, prefer checking the provider's current pricing page before quoting numbers externally.

Current Google image pricing notes

Assume 1024x1024 pricing unless noted otherwise.

Model Pricing
gemini-2.5-flash-image $0.039 per image
gemini-3.1-flash-image-preview $0.067 per image
gemini-3-pro-image-preview $0.134 per image

Current OpenAI image pricing notes

Assume medium quality and 1024x1024 pricing.

Model Pricing
gpt-image-1 $0.042 per image
gpt-image-1.5 $0.034 per image
gpt-image-1-mini $0.011 per image

These numbers were added from the providers' official pricing pages and should be treated as guidance, not permanent constants.

Provider Notes

OpenAI

  • Default model: gpt-image-1.5
  • Uses POST /v1/images/generations
  • With --ref, uses POST /v1/images/edits
  • Best when the user wants explicit image sizes like 1024x1024 or 1536x1024
  • Supports --size, --n, and OpenAI request fields such as --quality and --style
  • If the user gives only an aspect ratio, the script maps that to the closest supported OpenAI size
  • Reference images are supported in this skill for GPT Image models such as gpt-image-1.5, gpt-image-1, and gpt-image-1-mini

Google Gemini Image ("Nano Banana")

  • Default model: gemini-2.5-flash-image
  • Uses POST https://generativelanguage.googleapis.com/v1beta/models/<model>:generateContent
  • --ar is sent in generationConfig.imageConfig.aspectRatio
  • For --n > 1, the script makes sequential calls and asks for one distinct variation per call
  • --size is not used for Nano Banana in this skill
  • Reference images in this skill are supported for gemini-3.1-flash-image-preview and gemini-3-pro-image-preview

Output Behavior

  • If --n 1, --output is treated as the final file path; when no extension is provided, the script appends one
  • If --n > 1 and --output includes an extension, files are saved as name-1.png, name-2.png, etc.
  • If --n > 1 and --output has no extension, it is treated as a directory

Reference Images

Use --ref or --reference to pass one or more source images.

This makes the skill compatible with baoyu-xhs-images style reference-image chains, where image 1 is generated first and later slides reuse image 1 as --ref for visual consistency.

Example with OpenAI edits:

node ${SKILL_DIR}/scripts/main.mjs \
  --provider openai \
  --model gpt-image-1.5 \
  --prompt "Turn this product photo into a clean editorial hero image" \
  --ref source.png \
  --output edited.png

Example with Gemini reference-image prompting:

node ${SKILL_DIR}/scripts/main.mjs \
  --provider google \
  --model gemini-3.1-flash-image-preview \
  --prompt "Keep the subject identity but restyle this as a cozy watercolor poster" \
  --ref character.png \
  --output poster.png

Selection rules:

  1. For OpenAI + --ref, prefer GPT Image models such as gpt-image-1.5.
  2. For Google + --ref, prefer gemini-3.1-flash-image-preview or gemini-3-pro-image-preview.
  3. If the user wants consistent follow-on images from source art, use reference images instead of plain text-only prompting.

Interop

prompt-to-image interoperates cleanly with both baoyu-xhs-images and markdown-to-image.

With baoyu-xhs-images

Use prompt-to-image as the image generation engine behind the XHS workflow:

  1. baoyu-xhs-images creates the cover and content prompt files
  2. Generate image 1 without --ref
  3. Generate images 2+ with --ref <image-1> for visual consistency

This works because prompt-to-image supports:

  • --promptfiles
  • --ref
  • --ar 3:4
  • --quality 2k

For the OpenAI path, --quality 2k is normalized to an OpenAI-compatible high-quality request.

With markdown-to-image

Use prompt-to-image output as the visual layer for Inkframe:

  1. Generate the image with prompt-to-image
  2. Convert the saved image to a base64 data URL
  3. Put that data URL into Inkframe design fields such as backgroundImageUrl or contentBoxImageUrl
  4. Render the final card/poster with markdown-to-image

All three together

Recommended pipeline:

  1. baoyu-xhs-images handles Xiaohongshu-specific structure, style, and prompt planning
  2. prompt-to-image renders the actual images
  3. markdown-to-image composes branded summary cards, title cards, or text overlays using those generated images

Use this combined workflow when the user wants both AI-generated illustration and polished typographic composition.

Working Style

  1. Prefer concise but concrete prompts. Mention subject, style, composition, lighting, and mood.
  2. Use --size when the user gives an exact output size and you are on OpenAI.
  3. Use --ar when the user cares more about layout shape than exact pixel dimensions, especially on Gemini.
  4. For multiple concepts, keep --n modest unless the user explicitly wants a large batch.
  5. Save to a real file path and hand the user the created image file(s), not just console output.

Error Handling

  • Missing provider key: fail with a clear setup message
  • No image returned: surface the provider response details if available
  • Unsupported combination: prefer warning plus best-effort behavior, unless the request cannot work

Test Prompts

These are good manual eval prompts for this skill:

  1. Create two square app icon concepts for a budgeting app called PennyTrail. Make them warm, simple, and premium.
  2. Use Nano Banana to generate a vertical travel poster for Kyoto in rain, woodblock print style, lots of negative space.
  3. Generate an OpenAI image for a YouTube thumbnail about debugging React hydration issues. High contrast, modern, readable composition.
Weekly Installs
1
Repository
ericjy/inkframe
GitHub Stars
1
First Seen
5 days ago
Security Audits
Gen Agent Trust HubPass
SocketFail
SnykPass
Installed on
amp1
cline1
opencode1
cursor1
kimi-cli1
codex1