Article Illustrator

• GLM_API_KEY
• glm_api_key
• api_key

Inputs

• article: The full article text (markdown or plain text)
• image_count (optional): Number of images to generate, 2–5 (default: 3)
• orientation (optional): portrait (1088x1920, default) or landscape (1920x1088)
• language (required): Language for text in generated images — zh, en, ja, ko, fr, de, es. Always ask the user; never default or infer.

Workflow

Step 0: Check API Key

If ZAI_API_KEY is not configured (or compatible fallback key names are missing), stop and tell the user how to configure it. Check ~/.codex/.env first, then environment variables, then fallback config sources. Do not proceed without a valid key.

Step 1: Validate Input

Before generating any images:

• Confirm article is at least 200 words. If shorter, warn user: "Article may be too short for meaningful illustration — proceed anyway?"
• Count major sections or topic transitions (H2 headings, paragraph breaks). This determines insertion points.
• Ask the user which language to use for image text. Never infer from the article language.

Step 2: Generate Image Prompts

Read the scrapbook system prompt from references/scrapbook-prompt.md. If the user asks for a numbered style such as 风格1 or 风格3, also read references/style-library/STYLE_INDEX.md and match the requested style.

Using that prompt, analyze the article and output a JSON plan:

{
  "project_title": "Article title — Scrapbook Style",
  "style": "Physical Mixed-Media Scrapbook",
  "total_images": 3,
  "images": [
    {
      "image_id": 1,
      "title": "Image caption",
      "description": "300-500 character visual description in scrapbook style...",
      "insert_after": "Exact sentence or heading after which to insert the image"
    }
  ]
}

Target 1 image per 300–400 words of article content.

For WeChat/article visuals, prefer a poster-like composition over a dense dashboard:

• fewer text labels
• bigger typography
• one main idea per image
• avoid many tiny notes crammed into one frame
• avoid mixed Chinese/English labels unless absolutely necessary

The skill now supports numbered reference styles:

• 风格1 / style 1: Board System
• 风格2 / style 2: Blue Poster
• 风格3 / style 3: Character Sticker
• 风格4 / style 4: Editorial Story
• 风格5 / style 5: Bold Metaphor Scrapbook

Step 3: Generate All Images in Parallel

Launch all image generations simultaneously.

python3 scripts/generate.py "<description_1>" --language <lang> --size 1088x1920 &
python3 scripts/generate.py "<description_2>" --language <lang> --size 1088x1920 &
python3 scripts/generate.py "<description_3>" --language <lang> --size 1088x1920 &
wait

The bundled script always calls glm-image via Z.AI / BigModel-compatible API.

On failure: if generation fails for one image, log the error and continue with remaining images. Do not abort the entire run.

Step 4: Compose Final Markdown

Insert each image after its designated insert_after anchor:

![Image caption](image_url)

If the insert_after anchor is not found verbatim, insert at the nearest paragraph break.

Output Format

Return the complete markdown article with:

1. All original article text preserved exactly
2. Generated images inserted at logical break points
3. Captions from the JSON plan
4. A summary line at the end: <!-- Illustrated: N images generated, M failed, total cost: $X.XX -->

Edge Cases

• Short article (<200 words): Warn and confirm before proceeding
• Generation failure: Log error inline, continue with other images — do not abort
• Anchor not found: Insert at nearest paragraph break instead
• No API key: Surface error immediately with setup instructions
• All images fail: Return original article unchanged with error summary

Quality Gate

Reject and regenerate if any image has one of these problems:

• obvious typo or missing letters in large text
• too many tiny labels or fake microtext
• line-art / slide-deck / wireframe feel instead of physical scrapbook feel
• visual clutter that makes the main idea hard to grasp
• mixed-language labels that look accidental rather than intentional

For WeChat specifically:

• default to portrait unless the user explicitly wants cover or landscape
• keep image text short enough that each phrase is likely to render correctly