grainy-duotone

Imported from okou-ai/vm0-skills (illustration-template/grainy-duotone/SKILL.md). Install upstream with npx skills add okou-ai/vm0-skills --skill grainy-duotone. Copyright stays with the author.

/grainy-duotone — locked grainy two-tone editorial style

This style produces small editorial spot illustrations in a warm hand-drawn idiom: grainy charcoal-stippled black-ink figures (almost always a hand or pair of hands) interacting with free-floating flat color blocks, set on a warm cream background and dotted with scattered ambient marks. The base is locked; each invocation spins five creative dials.

Prompt interpretation

The user will usually give a short brief — sometimes a scene phrase ("hand raising a flag"), sometimes a metaphor + palette ("celebrating a launch, teal + yellow"), sometimes the full row of dials. Translate the brief into a complete scene without stopping to ask:

1. Pick a scene metaphor. Identify the verb the illustration is about (launch, grow, connect, discover, communicate, reflect, navigate, care, transform, build). If the user only gave a feeling, choose the most evocative everyday gesture that conveys it.
2. Pick a two-color palette. Two flat hues only. Lean cool for calm / reflective scenes, warm for celebratory / driven scenes.
3. Pick a complexity level (L1 / L2 / L3). L1 for a single quiet gesture, L2 for a balanced 1–2 prop scene, L3 for a layered editorial-collage piece.
4. Pick the mood. The mood should flow from posture + palette + density together; do not contradict it with a clashing color or busy density.
5. Honor the locked frame — cream background, grainy charcoal hands, outline-free color blocks, patterns layered on top, scattered ambient marks. Those never move.

Bias toward gestural compositions where a hand or pair of hands is the focal point and color blocks support the gesture (cuff, prop, surround). Avoid full bodies, faces, and crowded narrative scenes — this style is editorial spot illustration, not full scene illustration.

Locked style axes (NEVER vary)

Background

• Warm cream off-white — #F5F1EA or close — fills the canvas edge to edge
• No vignette, no edge shading, no gradient, no paper texture, no border, no frame
• The cream is unbroken negative space around the scene

Figures (hands, arms, faces, any organic body part)

• Drawn in grainy stippled charcoal-texture black ink — rough sketchy fill that reads as pencil/charcoal grain, not smooth flat black
• Tone variation lives inside the grain — lighter stipple for highlights, denser for shadows
• Contour lines are confident hand-drawn black ink, slight wobble allowed
• Never smooth vector fills, never crosshatched engraving, never airbrush
• Hands are the typical subject — palms, fists, grips, cupped shapes. Faces and full figures are rare.

Color blocks (the two-tone shapes)

• Two flat hues per piece plus black ink — no third color, no gradient, no shading
• Zero black outlines on color shapes — the shapes are free-floating, painted-first, with rough hand-painted edges (see feedback-blue-blocks-no-border rule from the broader spec)
• Ink lines may sit on top of a color block but must never trace its perimeter
• Color blocks are typically cuffs/sleeves under the hand, props the hand interacts with, or supporting silhouettes behind the gesture
• Color blocks are often a slightly irregular blob rather than a geometrically precise shape

Internal patterns (the texture inside color blocks)

• Every color block carries a pattern layered on top of its flat fill, drawn in black ink
• Pattern vocabulary: grid · dots · spiral · diagonal hatching · cross-hatching · wavy parallel lines · squiggles · concentric circles · tiled triangles · plus-signs · halftone dots
• One block typically carries one pattern; if two blocks coexist, give them different patterns (e.g. spiral vs. grid)
• Patterns are loose and slightly imperfect — not a CAD grid

Ambient marks (the scattered confetti)

• A handful of small black-ink marks float in the cream space around the scene
• Mark vocabulary: small zigzag · tiny triangle · x mark · small circle · short straight line · plus sign · dot · curly squiggle · dashes · chevron
• Counts: L1 ≈ 2–3 marks, L2 ≈ 5–7 marks, L3 ≈ 10–12 marks
• Marks are scattered, never aligned to a grid, never forming a recognizable pattern of their own

Composition

• Square 1:1, centered, with breathing room around the focal gesture
• One clear focal point (the hand or pair of hands) — never two competing subjects
• No text, no logos, no captions, no headlines, no borders

Variable axes (the five dials)

These are the only things that should change between pieces.

Brief template

Palette:    {two colors}
Elements:   {hand(s) + 1-3 props}
Scene:      {Launch | Grow | Connect | Discover | Communicate | Reflect | Navigate | Care | Transform | Build} — {one-line metaphor}
Complexity: {L1 minimal | L2 balanced | L3 rich}
Mood:       {Celebratory | Calm | Curious | Reflective | Determined | Warm | Quiet}

Generation guidance

This style was tuned with a multimodal image-edit model that accepts both a written prompt and a reference image set. Pass the seven reference PNGs in this folder as the style anchor — the prompt alone is not sufficient to keep the grain texture, outline-free color rule, and ambient-mark vocabulary aligned.

Model-specific note (fal-ai/nano-banana/edit, the model the style was tuned on):

• Always restate every locked-style rule in the prompt. The model does not retain locked rules across calls.
• Especially restate: (a) outline-free color blocks (b) cream background with no vignette (c) grainy stippled fill on the hand, not smooth black (d) two colors only, no third hue.
• Never write "match the reference grid" — the model interprets that as a multi-cell layout request. Write "match the reference style".
• For the ladder / climb scene archetype, explicitly state that the ladder is rendered in plain black ink lines, NOT colored — otherwise the model colors the ladder.

If you are running a different multimodal edit model, keep the same semantic instructions: rules first, then scene, then references. Square 1:1 output.

Reference images (locked anchors)

These seven PNGs are the style anchors and should be passed as image references on every generation. They are NOT user-supplied source material — they are codified outputs from the framework's tuning round.

Evaluation checklist

After generation, the piece should pass:

• Background is warm cream — no white edge halo, no gradient.
• Hand(s) and any skin are grainy charcoal stipple — not smooth flat black.
• Color blocks have zero black outlines tracing their perimeter.
• At least one pattern (grid / dots / spiral / hatching / squiggles / wavy lines) sits inside each color block.
• Exactly two color hues + black are present — no third color.
• Ambient marks count matches the complexity level.
• One clear focal gesture — no competing subjects.
• No text, no logo, no border.

If any item fails, regenerate with the failed rule restated more emphatically in the prompt.