marsonshine/ai-skills/qwen-image-generator

Qwen Image Generator

Turn rough image requests into saved local files. Default to doing the work end to end: load preferences, ask only for still-missing settings, write one concrete English prompt, generate sequentially, save immediately, and report the file paths plus the final prompt used.

Compatibility

Works best with PowerShell, Node.js 22+, environment variables, and file tools. The PowerShell entry is thin; scripts/generate_qwen_image.ts owns the DashScope call, polling, downloads, and JSON output.

Runtime layout

Credentials

Resolve credentials in this order:

1. QWEN_IMAGE_API_KEY
2. DASHSCOPE_API_KEY

Optional: DASHSCOPE_BASE_URL

If no key exists, prepare the prompt and effective settings, report the missing credential, and ask the user to configure it locally. Do not submit generation or claim success until a key is available.

Preferences

Check for EXTEND.md in this order:

1. .baoyu-skills/qwen-image-generator/EXTEND.md
2. $XDG_CONFIG_HOME/baoyu-skills/qwen-image-generator/EXTEND.md
3. $HOME/.baoyu-skills/qwen-image-generator/EXTEND.md

If none exists, use the built-in defaults below and continue. Read references/config/first-time-setup.md only when the user wants to save or change persistent defaults; an image request does not require a preference file.

Built-in defaults when neither the request nor EXTEND.md overrides them:

• model: qwen-image-2.0
• style: flat-illustration
• size: 1024x1024
• text policy: avoid
• prompt enhancement: true
• watermark: false
• output directory: ./generated-images/

Working rules

1. Parse the request for subject, style, size, visible text, and output path.
2. Fill missing values from EXTEND.md, then built-in defaults. Explicit request values take precedence.
3. Ask only for still-missing values that materially change the result.
4. Write the final prompt in English.
5. Prefer project batch wrappers when they already handle reruns or selective regeneration. Otherwise run scripts/generate_qwen_image.ps1.
6. Generate sequentially by default. Only parallelize when the user explicitly asks.
7. Save each successful result immediately and report the saved path, final prompt, and effective settings.

Load references/prompting.md when the request needs preset choice, educational-image rules, or prompt-shaping guidance.