visual-atelier

name	visual-atelier
description	Use when the user asks for style-based image generation, image sets, social cards, carousels, posters, covers, or a specified number of generated images from given content, style, and dimensions. Supports 指定风格, 指定尺寸, 指定数量, 封面图, 轮播图, 海报卡片, 小红书图片笔记, and provider-native image generation with ZenMux or imagegen.

Visual Atelier

Turn source material or direct image requirements into a refined provider-native PNG image set. The skill supports general requests such as "use this style, this size, this content, make N images"; Xiaohongshu notes are one supported social-card use case, not the skill boundary.

Core Contract

• Do not generate style demo images by default. Use the static high-fidelity style examples as the style selection surface: vertical examples in assets/style-examples/, and 4:3 landscape examples in assets/style-examples-horizontal/.
• Show the static style examples first when the user has not already specified a style. Ask the user to choose by number or slug before generating final images.
• If the user already specifies a style, skip the style menu and proceed with that style.
• If the user gives a custom style not covered by the examples, write a concise style brief and ask only if the style is ambiguous.
• Support arbitrary final image dimensions and counts. For Xiaohongshu image notes, default to 1080 x 1350 and 4:5 vertical when the user does not specify size. For 4:3 landscape requests, use 1280 x 960 unless the user provides an exact size.
• Support image provider selection. Use zenmux only when ZENMUX_API_KEY is present and scripts/zenmux_generate_image.py --check passes; otherwise use built-in imagegen, unless the user explicitly chooses a provider.
• Never paste or store API keys in output files. Read keys from environment variables only.
• For ZenMux, default to base URL https://zenmux.ai/api/v1 and model openai/gpt-image-2, unless ZENMUX_BASE_URL, ZENMUX_IMAGE_MODEL, or ZENMUX_IMAGE_SIZE override the defaults. The bundled script defaults to REST calls, matching the oil-cover skill; the OpenAI SDK-compatible client is optional via --provider-client openai.
• Generate final images with the selected image provider by default. Do not use local HTML/CSS/Playwright rendering unless the user explicitly requests exact-text screenshots and accepts the possible visual downgrade.
• Keep one visual system across the final image set. Do not mix multiple built-in styles in one set unless the user explicitly asks for mixed styles.
• Create style-lock.md and visual-system.md before final generation. They should be based on the selected static style example or the user's custom style brief.
• For Xiaohongshu notes, create note-package.md and usually post-copy.md. For general image sets, create image-set-package.md instead.

Workflow

1. Load references only as needed.

   • Read references/style-presets.md before showing style options or applying a selected style.
   • Read references/image-providers.md before generating final images, especially when ZenMux is configured or requested.
   • Read references/article-parsing.md when the source is long, messy, multi-format, or not already split into image roles.
   • Read references/style-lock.md and references/visual-system.md after style selection and before final generation.
   • Read references/cover-design.md for Xiaohongshu covers or any request that needs a first/cover image.
   • Read references/publishing-copy.md only when the output includes Xiaohongshu publishing copy.
   • Read references/rendering-and-qa.md before validating final assets.

2. Parse the user's image request.

   • Extract: purpose/platform, source content, target audience, image count, dimensions, aspect ratio, style, provider, output folder, and whether publishing copy is needed.
   • If the request is a Xiaohongshu note and count is unspecified, choose 5-8 pages for most knowledge/tutorial notes, 3-5 for quick tips, and 8-12 for full guides.
   • If the request is a general image set and count or dimensions are missing, infer only when the intent is obvious; otherwise ask a short clarifying question.
   • Preserve claims. Do not invent metrics, testimonials, product facts, logos, screenshots, or official endorsements.

3. Show static style examples instead of generating demos.

   • Match the example set to the requested aspect ratio. Use assets/style-examples-horizontal/ for 4:3 landscape requests; otherwise use assets/style-examples/.
   • Vertical examples in assets/style-examples/:
     • style-01-clean-lifestyle-guide.png
     • style-02-signal-playbook.png
     • style-03-editorial-magazine.png
     • style-04-paper-system-mirror.png
     • style-05-handdrawn-notebook.png
     • style-06-ai-tool-lab.png
     • style-07-daily-lime-lab.png
     • style-08-obsidian-neon-knowledge.png
   • 4:3 landscape examples in assets/style-examples-horizontal/:
     • style-01-clean-lifestyle-guide-4x3.png
     • style-02-signal-playbook-4x3.png
     • style-03-editorial-magazine-4x3.png
     • style-04-paper-system-mirror-4x3.png
     • style-05-handdrawn-notebook-4x3.png
     • style-06-ai-tool-lab-4x3.png
     • style-07-daily-lime-lab-4x3.png
     • style-08-obsidian-neon-knowledge-4x3.png
   • Present them as existing high-fidelity examples. Do not call any image provider during style selection.
   • Do not create demo-overview.png, contact sheets, generated demo covers, or per-request style previews unless the user explicitly asks for new explorations.
   • After showing the examples, stop and ask the user to choose a style if no style was specified.

4. Lock the selected style.

   • Create style-lock.md in the output folder or selected-style subfolder.
   • Record selected style slug, selected style example path or custom style brief, visual promise, locked traits, page/image inheritance rules, and forbidden drift.
   • Create visual-system.md from the style lock. Translate the selected style into typography mood, color ratios, layout rhythm, component shapes, background treatment, graphic devices, and footer/label rules.

5. Plan the final image set.

   • For Xiaohongshu notes: plan cover, body pages, and CTA/summary page; write short in-image text and exact companion copy.
   • For general image sets: plan one role per image, exact dimensions, required text, subject matter, composition, and any sequence logic.
   • Use short in-image text. Put exact long copy in note-package.md, post-copy.md, or image-set-package.md.

6. Generate final images with the selected provider.

   • Use the selected provider for every final image.
   • Before choosing ZenMux, run python3 scripts/zenmux_generate_image.py --check. If the default REST check passes, ZenMux is locally configured.
   • If ZenMux /images/edits returns 403 Forbidden, let the script fall back to /images/generations before falling back to imagegen. If ZenMux generation also returns 403 Forbidden, record the failure in generation-record.md and fall back to imagegen unless the user explicitly requires ZenMux.
   • If using ZenMux, prefer scripts/zenmux_generate_image.py --image assets/style-examples*/style-0N-*.png ... when the static style example can help preserve style. For 4:3 landscape work, prefer the matching horizontal example. Add user-provided reference images with additional --image arguments when relevant.
   • If using imagegen, carry the selected style example's visual traits into the prompt.
   • For Xiaohongshu notes, export as 01-cover.png, 02-{role}.png, etc.
   • For general image sets, export as 01-{role}.png, 02-{role}.png, etc.
   • Regenerate images that fail dimensions, major content requirements, style consistency, readability, or claim safety.

7. Write the companion document.

   • Xiaohongshu note: save note-package.md; create post-copy.md when publication copy should be separate.
   • General image set: save image-set-package.md with original request, selected style, dimensions, count, per-image role, prompt summary, text used, output list, and provider/model.
   • Always create generation-record.md with provider, model, dimensions, selected style, reference images used, and any fallback or regeneration notes. Never include secrets.

8. Validate before delivery.

   • Validate static style examples after skill changes:

     python3 scripts/validate_package.py assets/style-examples --style-examples-set
     

   • Validate static 4:3 landscape examples after skill changes:

     python3 scripts/validate_package.py assets/style-examples-horizontal --horizontal-style-examples-set
     

   • Validate Xiaohongshu final packages:

     python3 scripts/validate_package.py <output-dir> --width 1080 --height 1350
     

   • Validate general image sets:

     python3 scripts/validate_package.py <output-dir> --general-image-set --width {width} --height {height}
     

   • Inspect final images for text overflow, blank lower halves, style drift, unsupported claims, fake logos, watermarks, template sameness, and incorrect dimensions.

Default Output Structures

For Xiaohongshu notes:

outputs/xhs-note-{YYYY-MM-DD}-{slug}/
  {selected-style}/
    style-lock.md
    visual-system.md
    01-cover.png
    02-{role}.png
    ...
  note-package.md
  post-copy.md
  generation-record.md

For general image sets:

outputs/image-set-{YYYY-MM-DD}-{slug}/
  {selected-style}/
    style-lock.md
    visual-system.md
    01-{role}.png
    02-{role}.png
    ...
  image-set-package.md
  generation-record.md

If the user specifies an output folder, use it.

Style Selection Response

When the user has not chosen a style, show the static examples with image links and short guidance:

下面是内置高保真风格样张，请选编号或 slug：

1. clean-lifestyle-guide：适合效率、学习、清单、轻教程。
   ![clean-lifestyle-guide](/absolute/path/assets/style-examples/style-01-clean-lifestyle-guide.png)
2. signal-playbook：适合技术流程、排错、部署、行动手册。
   ![signal-playbook](/absolute/path/assets/style-examples/style-02-signal-playbook.png)
3. editorial-magazine：适合深度思考、产品分析、高级感内容。
   ![editorial-magazine](/absolute/path/assets/style-examples/style-03-editorial-magazine.png)
4. paper-system-mirror：适合 AI 提示词、工作流复盘、内容系统、创作者方法论。
   ![paper-system-mirror](/absolute/path/assets/style-examples/style-04-paper-system-mirror.png)
5. handdrawn-notebook：适合新手科普、学习笔记、心智模型。
   ![handdrawn-notebook](/absolute/path/assets/style-examples/style-05-handdrawn-notebook.png)
6. ai-tool-lab：适合 Codex、Agent、MCP、Skills、AI 工作流。
   ![ai-tool-lab](/absolute/path/assets/style-examples/style-06-ai-tool-lab.png)
7. daily-lime-lab：适合 AI Skill 推荐、工具日报、前端/设计资源推荐、开源项目介绍、插件/模板种草。
   ![daily-lime-lab](/absolute/path/assets/style-examples/style-07-daily-lime-lab.png)
8. obsidian-neon-knowledge：适合 Obsidian、Codex、知识库、AI 工作流、系统搭建、工具生态。
   ![obsidian-neon-knowledge](/absolute/path/assets/style-examples/style-08-obsidian-neon-knowledge.png)

请选择风格编号或 slug，我再按你的内容、尺寸和数量生成最终图片。

Use absolute local paths in Markdown image tags so the app can render them. For 4:3 landscape requests, use the matching files under assets/style-examples-horizontal/ in this same response pattern.

Final Image Prompt Shape

Use this prompt shape for each final image:

Use case: {social-card / infographic-diagram / poster / product-card / custom}
Asset type: provider-native PNG image {N} of {total}
Provider: {imagegen / zenmux}
Dimensions: {width} x {height}
Selected style: {style slug or custom style name}
Style reference: {static example path or custom style brief}
Style lock: {3-6 required visual traits from style-lock.md}
Visual system: {5-8 executable rules from visual-system.md}
Image role: {cover / framework / steps / comparison / checklist / CTA / custom role}
Content source: {short factual summary}
Short in-image text:
Title: "{title}"
Subtitle: "{subtitle}"
Bullets or labels: "{item 1}" / "{item 2}" / "{item 3}"
Composition: {specific layout instructions from selected style and requested aspect ratio, not a generic template}
Craft requirement: preserve style-level typography care, color richness, component shapes, decoration density, and readable hierarchy.
Text handling: use short phrases only; exact long wording goes in the companion Markdown document.
Constraints: provider-native final image, no watermark, no fake data, no unsupported claims, no fake logos, no palette-only style inheritance, no HTML/CSS template look, no stretching a vertical layout into a horizontal canvas.

If the provider does not preserve exact text, shorten the in-image wording and regenerate. Do not switch to HTML/CSS just to repair small text errors unless the user explicitly requested exact-text screenshots.

Handoff

Final responses should list:

• Selected style.
• Image provider and model.
• Output folder.
• Final dimensions and image count.
• Final PNG paths.
• style-lock.md path.
• visual-system.md path.
• Companion Markdown path.
• Recommended title and publishing hook when it is a Xiaohongshu note.
• Validation result.