drawing-bak（绘图 bak）

When to Use

User needs original AI drawings, coloring pages, or simple educational illustrations, especially for children. Agent turns vague requests into model-portable prompts, age-appropriate style choices, and print-ready constraints that work in OpenClaw or any image model the user already has.

Use this when the real problem is not "pick the best model" but "get a clean result fast": full color vs coloring page, preschool vs older kids, one-off scene vs consistent series, or first prompt vs refinement loop.

Architecture

Memory lives in ~/drawing/. If ~/drawing/ does not exist, run setup.md. See memory-template.md for structure and status fields.

~/drawing/
|- memory.md            # Default age bands, style preferences, and output habits
|- winning-prompts.md   # Prompts that already worked well for this user
|- style-notes.md       # Preferred palettes, line weight, and recurring motifs
`- series.md            # Character and scene anchors for multi-page sets

Quick Reference

Load only the smallest file needed for the current bottleneck.

D.R.A.W. Protocol

Use this four-step protocol before writing a prompt.

Decide the output mode

Pick one mode first and make it explicit:

Reduce the scene

Keep 1-3 focal elements. If the user wants a coloring page, simplify harder than feels natural. A "cute but detailed" prompt usually fails for children because the model fills the page with tiny shapes and noisy backgrounds.

Anchor the style

Pick one style pack and one composition. Do not stack "watercolor + kawaii + paper cut + cinematic" in the same prompt. The style must support the use case:

• soft color illustration for story or gift
• bold cartoon for fast recognizability
• black outline sheet for coloring
• classroom visual for learning

Wrap with constraints

End with non-negotiables:

• age appropriateness
• original characters only unless the user explicitly asks otherwise
• no scary mood, no violence, no text unless essential
• no watermark, no logo, no cropped subject
• print target if relevant: A4 or US Letter, portrait or landscape

For full templates and slot order, load prompt-system.md.

Fast Prompt Starters

Full-color drawing

Create an original children's illustration for ages [age range].
Subject: [main subject].
Scene: [simple setting with 1-2 supporting elements].
Style: [style pack].
Composition: [portrait or landscape], centered focal point, generous breathing room.
Color: [palette or mood].
Keep it friendly, clear, and easy to recognize. No text, no watermark, no scary details.

Coloring page

Create a printable black-and-white coloring page for ages [age range].
Subject: [main subject].
Use thick clean outlines, large closed shapes, minimal background, white page, and no shading.
Keep only the essential elements needed to recognize the scene.
No text, no gray fill, no tiny decorations, no cropped objects, no watermark.

Educational drawing

Create a simple educational illustration for children learning [topic].
Show [concept] clearly using one main scene and only the supporting objects that help explain it.
Use [style pack]. Keep labels out of the image unless the user specifically needs them.
Make it accurate, friendly, uncluttered, and easy to print.

Core Rules

1. Start with the child, not the model

• Ask or infer the age band, use case, and whether the result is for coloring, printing, or screen viewing.
• Preschool output should be radically simpler than "pretty" output for older kids.
• If the user does not care about the model, optimize for prompt quality and let OpenClaw use the configured image model.

2. Separate subject, style, and constraints

• Prompts work better when the main subject, scene, style, and restrictions are stated in distinct blocks.
• Avoid vague adjectives like "nice", "professional", or "beautiful" unless you translate them into visual traits.
• Reusable prompts should keep variables obvious: subject, age, style, color mode, page format.

3. Match detail level to coloring difficulty

• Coloring pages for ages 3-5 need big shapes, thick outlines, and almost no background clutter.
• Ages 5-7 can handle one or two supporting objects and mild patterning.
• Older kids can tolerate moderate detail, but the subject must still read instantly.
• If the user says "for kids" and gives no age, default to the safer 5-7 profile.

4. Lock invariants for any series

• For multiple pages, repeat the same character traits, line style, camera distance, palette family, and page framing.
• Preserve one prompt block called "always keep" and reuse it unchanged across generations.
• Change one variable at a time between attempts so the cause of improvement stays visible.

5. Design for print when print matters

• Explicitly request white background, clean margins, uncropped subject, and portrait or landscape orientation.
• Coloring pages should be ink-friendly: no halftones, no gray textures, no faux paper grain.
• If the result will be printed at home, prefer centered compositions and avoid edge-to-edge details.

6. Run a tight validation loop

• Generate a small batch, inspect the actual result, then change only one thing: style, detail, composition, or negative constraints.
• If hands, faces, or proportions fail, simplify the pose before adding more detail.
• If the background is noisy, remove it rather than trying to "clean it up" with more adjectives.

7. Keep the output kid-safe and IP-safe by default

• Default to friendly expressions, calm scenes, and original characters.
• Avoid brand characters, copyrighted mascots, or lookalikes unless the user accepts that risk.
• Do not use real child photos as references unless the user explicitly wants that workflow and trusts the selected provider.

Common Traps

• Asking for "cute and detailed" coloring pages -> the model adds clutter that frustrates children.
• Mixing too many styles in one prompt -> inconsistent output and lower controllability.
• Forgetting the age band -> detail level drifts and the result misses the user.
• Leaving text inside the image -> most image models still mangle labels in children's material.
• Changing five things at once -> you cannot tell which edit improved the result.
• Using the same prompt for color and coloring-page modes -> line quality and composition drift immediately.

Security & Privacy

Data that leaves your machine:

• Prompt text and any reference images sent to the selected image provider through OpenClaw or another image client.

Data that stays local:

• Style preferences, winning prompts, and recurring constraints under ~/drawing/.

This skill does NOT:

• Force a specific model purchase or provider.
• Upload files unless the active image workflow already does so.
• Store real child photos by default.
• Modify its own skill file.

Trust

By using image generation, prompt text and optional reference images may be sent to third-party model providers. Only use providers you trust, and avoid sending unnecessary personal details or identifiable child photos when a generic description would work.

Related Skills

Install with clawhub install <slug> if user confirms:

• art - Broader art direction, critique, and technique guidance beyond image prompting.
• logo - Prompt patterns for icons, marks, and other cleaner graphic outputs.
• design - Clarify visual taste and creative direction before locking a drawing style.
• graphic-design - Improve layout, print choices, and supporting visual materials around the drawing.

Feedback

• If useful: clawhub star drawing
• Stay updated: clawhub sync