Restrained Hand-Drawn Illustration

Create an original small, playful illustration in the visual system defined in style-system.md. Treat the supplied references as constraints, not source material to copy: never copy distinctive characters, exact compositions, logos, or readable text.

Route the request

Classify each supplied image before making anything:

Choose one route:

1. Generate — brief or concept to illustration.
2. Photo/image transform — redraw a supplied target in the system while retaining the agreed essentials.
3. Reference analysis — report fixed rules, variables, and a reusable prompt; do not generate unless asked.
4. Prompt-only — return the compiled prompt, production recipe, and output specification only.
5. Animate — make a short looping GIF from a generated or supplied still, selecting motion from motion-system.md.

Ask one concise question only if the missing information materially changes the output: whether a supplied image is a target or reference, required ratio when no source exists, exact handwritten text, or requested animation intent. Otherwise use the defaults below.

Defaults and output contract

• Preserve the supplied target image's width:height ratio exactly. If there is no target, default to a 1:1 square.
• Render at least 1536 px on the long edge (or the user's requested dimensions); the drawing normally occupies only 20–45% of the canvas and leaves generous empty space.
• Deliver the final still as an RGB JPG. Convert losslessly generated PNG/WebP results with scripts/export_raster.py only after visual approval.
• For animation, deliver a looping GIF at the still's aspect ratio, normally 3–5 seconds, 8–12 fps, 12–36 frames. Use only a small, meaningful loop rather than continuous busy motion.
• If lettering is requested, keep it short (normally 1–6 words). Handwrite it as a drawn shape; do not rely on an invented font or claim exact spelling when it is not legible. For brand-critical or longer text, typeset/letter it in a separate post-processing pass and confirm the final spelling.

Production workflow

1. Find the tiny joke. Reduce the brief to one small visual surprise: an object doing one unexpected thing, a tiny character reacting, or one playful exchange. Limit the cast to 1–2 subjects and 0–2 supporting marks. Do not make an establishing shot.
2. Lock the scale. Read style-system.md. Place the vignette inside a compact visual footprint with 55–80% of the canvas blank. A requested large object is still drawn simply; do not turn it into a grand scene, poster, cityscape, or technical diagram unless the user explicitly overrides this rule.
3. Select the line profile. Read line-profiles.md. Choose one profile rather than averaging the references. Default to soft-crayon-outline for character/object vignettes; use fine-gesture only when the brief calls for airy, delicate marks. State the profile explicitly in the prompt.
4. Select the recipe. Build the form from a few broad, tactile, intentionally open contour strokes plus a smaller number of thin gesture marks. Each visible stroke needs a natural pressure path: light start, fuller middle, eased or lifted end. Leave 1–3 meaningful breaks where an edge turns away, overlaps another form, or does not need explanation; never seal every object into a complete outline. Add one muted flat accent only when it carries the joke or makes the subject readable.
5. Compile the prompt. Use prompt-compiler.md. State the required aspect ratio, compact subject footprint, selected line profile, and prohibited scene scale. For an edit target, state what must be retained and what may be simplified.
6. Generate and inspect. Use the image-generation capability with the actual edit target attached when applicable. Inspect at full frame: tactile line-weight hierarchy, soft but decisive contours, visual smile/curiosity, sparse color, small vignette scale, large pale negative space, no accidental lettering, and no copied reference residue. Regenerate or revise when any gate fails.
7. Finish still output. Convert the approved result to JPG with scripts/export_raster.py jpg INPUT OUTPUT. Verify the output dimensions match the intended ratio.
8. Animate only when requested. Define a tiny loopable motion in words, then create consistent frames or controlled layer movement. Read motion-system.md, inspect beginning/end continuity, and compile with scripts/export_raster.py gif. Do not animate every element.

Quality gate

Reject and retry when the result has any of these: poster-scale or full-frame subject; realistic city/landscape; cinematic establishing shot; technical rendering; a single uniform line width; blunt cut-off stroke ends; fully sealed contour on every form; hairline sketch marks used for the silhouette; scratchy repeated outlines; hard geometric joins; polished vector uniformity; thick cartoon outlines with perfectly smooth edges; mechanically regular spokes/repetition; exact radial symmetry; more than two color accents; saturated/neon color; dark or busy background; realistic shading; dense scenery; decorative clutter; copied wording, logo, character, or composition; rigid symmetry without a small joke; unreadable required text; or motion that jitters, flickers, or does not loop cleanly.

Before delivery, state: input role, retained elements (when transforming), chosen palette, output dimensions/ratio, and—if applicable—the GIF loop action and duration. Attach only the requested final format(s), not intermediate explorations unless asked.