Collect a character trait card. Return JSON {"name":"...","silhouette":"...","signatureDetails":"...","palette":"...","personality":"..."}. silhouette - how a character is recognized by one silhouette (shape, not 'person'); signatureDetails — 2-3 branded item parts; palette – 3 basic colors + accent; personality - character in 1-2 words, reflected in appearance.
Mechanics: the generator assembles a character from the characteristics you named; He thinks out everything that is not fixed anew at each generation - from here he “floats.” Recognition rests not on the name, but on a set of stable visual anchors. The main thing is the silhouette: if a character is recognizable by a black fill without details, he will be recognized in any scene. Then 2-3 signature details (scar, half-glasses, always a red scarf) - the brain catches them first. Limit the palette: 3 basic colors plus 1 accent; more - the image crumbles. Insight: describe features with object nouns, not mood adjectives ('triangular hood, brass brooch', not 'mysterious look') - the model draws objects, not abstractions. Second insight: add one contrasting 'impossible' element - it makes the character memeable and takes away from the stock face. A typical beginner's mistake: a bunch of small details without hierarchy - the model chooses different ones each time, and the character is not repeated. Silhouette and palette first, details later.