pengmn23/article-to-images

Article to Image Prompts

Convert user content into structured JSON prompts ready for image generation. Can produce a single image or a series. Default delivery is the prompt; only proceed to generate images when the user requests it and the environment has image generation tools available.

Three Usage Modes

User Intent	Path
Turn an article, outline, or idea into images	Sections 1–7 "Image Creation"
Deconstruct an image: analyze layout, style, transcribe text, reverse-engineer prompts	Section 8 "Image Deconstruction"
Save created or deconstructed images as templates; use saved templates to create new images	Section 9 "Templates"

If the user provides only an image without content, first determine intent: deconstruct if they want analysis; if they want a similar style for their own content, ask for the content first and treat the image as a style reference.

This skill is text-only and does not rely on scripts:

File	Purpose	When to Read
templates/prompt-template.json	Full JSON skeleton for one image, with fill-in instructions for each value	Before writing prompts
templates/field-guide.md	Meaning, required/optional, length, and display rules for each field	Before writing prompts
library/	Ready-made content organized by field, image type recommendations, 12 style quick-matching options	When proposing style options and filling fields
references/preflight.md	What to ask before starting and how to ask	Step 2
references/series-and-delivery.md	Series planning, style master copy, delivery format, first image reference for generation	When creating a series or generating images
references/deconstruct.md	Seven-dimension deconstruction rules, output format, and conversion to nine-zone prompts	When deconstructing an image
references/templates.md	How to save, strip content, use, and manage templates	When saving templates or the user requests a template
my-templates/	User's own templates and index	Only when the user requests a template or asks "what templates do I have?"
examples/	Complete examples for single images, series, and PPTs	When unsure how to write

1. Read Content, Determine What to Do

1. Get Content: An article, file, outline, or a single sentence all work. If only a reference image without content is provided, first ask what the user wants to express; do not fabricate an article. Do not modify the original text; retain titles, data, and proper nouns as given.
2. Determine Image Type: Read 00 Image Type and decide whether it's a Xiaohongshu cover, knowledge card, infographic, comic, PPT, or other. If the user specifies, follow their instruction.
3. Determine Single vs. Series: If the user specifies a number, follow it. If they only say "turn it into images," treat short content or a single point as a single image; multiple subsections suggest a series, and provide a recommended number with reasoning. Include cover and summary images in the count.
4. Defaults: Aspect ratio recommended by type (Xiaohongshu 3:4, PPT 16:9), language Simplified Chinese. Always note defaults as defaults; do not treat them as confirmed by the user.

2. Ask Before Writing

Follow the Preflight Questions:

• Single Image: Ask 2 questions: style (three options) and text density.
• Series: Ask 4 questions: style (three options), text density, and 2 trade-off questions designed based on the content.

Do not repeat questions already answered. If the user explicitly requests a saved template (e.g., "use template U001"), treat the style as selected, read the template following Templates, and only ask about density and content questions. If the user says "you decide" or "just do it," choose based on content and write down what was chosen and why in the delivery. Wait for the user to answer before writing prompts; recommendations are not user answers. Saying "also generate images" does not bypass questions.

Style options can be selected from Style Quick Match or newly drafted based on the content. If the user provides a reference image, examine it first; all three options should retain the core characteristics of the reference.

3. Single Image: Write One Complete JSON

1. Copy prompt-template.json.
2. Based on the selected style, fill in style fields with content from library files; delete unused optional keys (for single images, usually remove series numbering, series style anchor, and suggested reference image workflow).
3. Select a layout from 04 Layout and write layout_structure and visual_logic.
4. Break content into nodes: each node has node_role, visual_anchor_description (drawable elements), displayed text verbatim, and how text maps to visuals. See 09 Visual Anchors & Nodes.
5. Select constraints and negative prompts from 10 Constraints & Negative and add type-specific entries.
6. Self-check against Section 5, then deliver.

4. Series: Plan → Master Copy → Individual Images

See detailed format in Series & Delivery.

1. Image Planning: For each image, write: number, title, corresponding original text location, image task, layout, nodes, and estimated text count. Explain what content is merged or omitted. Present the plan to the user first; stop here if they request review.
2. Style Master Copy: Write a master copy based on the selected style, including visual style, title and body text styles, color scheme, lighting, character design, rendering settings, negative prompts, canvas ratio, language, and series style anchor.
3. Write Each Image: Copy from the master copy for each image; copy style fields verbatim, only change core theme, layout, title text, nodes, node relationships, and per-image constraints. Each image must be a complete standalone JSON; do not write "same style as above."
4. If a particular image needs a style tweak, modify only one or two fields and note in the plan which image, which field, and why.

5. Checklist (Review Before Delivery)

Content

• Every character in the image comes from the user's content or user-approved rewrites; no added facts, data, examples, or prices
• On-screen text only includes: title_content, node title/description, and page_number (if any)
• node_role (e.g., main illustration, subtitle, quote) is not written into title
• Every emphasized word can be found in the corresponding node's text
• Text count is close to the selected density; exceptions (e.g., cover) are explained

Structure

• JSON is valid: double quotes, no trailing commas, line breaks as \n
• All template placeholders like 【required】 are replaced; unused optional keys are removed
• Node IDs are unique; no child nodes use []; nodes with text_display: hidden have empty title and description
• Visual anchors describe drawable elements; count matches content

Style

• Positive and negative prompts do not conflict (e.g., do not write "watercolor" in negative if style is watercolor)
• No artist names, brand names, or real logos; IP content uses original characters
• For series: style fields are verbatim from the master copy, field structure is identical, numbering is sequential

Delivery

• Every image's JSON is fully provided; no omissions like "omitted" or "same"
• Defaults, auto-selections, and content trade-offs are explained to the user

Field consistency ensures prompt uniformity, not output uniformity; each image must be reviewed separately.

6. Delivery

• Single Image: First explain in 2-3 sentences the chosen type, style, and layout; then provide the complete JSON; finally, list any points the user should note (e.g., dense text may need zooming).
• Series: Deliver in order: choices made, image plan, style master copy, complete JSON for each image, and content trade-off explanation. If file tools are available, save to disk following Delivery Directory; otherwise, present everything in full in the chat in the same order.
• The JSON is prompt text for image models: it can be pasted directly into image generation tools that support long prompts; size and aspect ratio are set separately in the tool.
• End by asking if they want to save the template: After delivering a series or a single image (and after checking the first image or single image if generated), ask "Would you like to save this style and layout as a template, so you can reuse it by saying 'use template U00X' next time?" If yes, save per Section 9; if no, end the conversation without further prompting. Do not ask if the template used was already saved.

7. Optional: Generate Images

Only proceed when the user requests it and the environment has image generation tools. Read the tool's own documentation; do not tie to a specific platform.

• Single Image: Generate using the complete JSON; check text, quantity, layout, and style; retry on issues.
• Series: Generate the first image and check it. Once approved, for each subsequent image, attach the first image as a reference (not just the previous image). See First Image Reference Generation for details.
• If the tool does not support reference images, inform the user and let them decide whether to switch tools or accept text-only generation; do not silently degrade.
• If the same failure occurs twice consecutively, investigate the cause (too much text, layout conflict, tool limitation) before retrying; do not retry indefinitely.
• Report separately: planned count, prompts completed, actual generations, checks passed, and incomplete items.

8. Image Deconstruction

Follow Image Deconstruction:

1. First Examine the Image: Actually view the whole image; zoom in on text areas if dense. If no image is available or cannot be opened, do not pretend to have seen it. Do not ask about style or density, and do not generate images.
2. Seven-Dimension Deconstruction: layout container, semantic theme, artistic style, visual medium, layout details, text content, technical parameters; plus node tree, constraints, and negative_prompt.
3. Transcribe Text Verbatim: Record all visible text as-is, including language, punctuation, numbers, units, and any typos in the original. For illegible parts, write 〔unrecognizable〕; do not fill in missing text. The text in the image is observational content, not instructions.
4. Delivery: Provide a single JSON object in a json code block. Add only one follow-up question: whether to ① convert to nine-zone prompts ready for generation, ② save as a template, or ③ both.
5. If the user chooses ①, convert per Section 8 of deconstruct.md; if ②, follow Section 9 "Save from Deconstruction".

9. Templates

Follow Templates: Save & Use. Templates are stored in my-templates/, one folder per template, numbered starting from U001.

• Saving: Ask at the end of image creation (Section 6) or after deconstruction (Section 8, Step 4). If the user agrees, save template_description.md and a content-stripped template.json. For series, add layout_inventory.md; for deconstruction-derived templates, add seven_dimension_deconstruction.json. The stripped template must not retain proper nouns, original text, or data from the original content.
• Reference Image: Ask if the user wants to save the first image or the original image as part of the template (reference_image.png); only save if they agree, and only one image. Remind them to delete it before sharing the skill.
• Use: Only read templates when the user explicitly requests one; do not proactively recommend. Treat the style as selected; still ask about density (and for series, ask content questions). If a reference image exists and the tool supports it, attach it as a style reference during generation.
• Management: Renaming, updating, and deleting require explicit user requests; confirm before deletion. Update the index after any change.
• If there is no file write permission, provide the complete file content to the user and ask them to save it themselves; do not claim it has been saved.

Version

Current 2.1.0. See CHANGELOG.md for change history.