Soft-Focus Color Haze

Create a fresh image in this template's visual language. Keep both retained reference files unchanged, and never use either reference's exact spatial arrangement or depicted subject as the default output.

Workflow

1. Read artifact-template.json and resolve its canonical background reference relative to this skill directory. Also resolve assets/subject-reference.png, the dedicated subject-led style reference.
2. Determine the mode from the user's brief and supplied images:
   • Background-only: no person, object, product, UI, typography, or other focal element is requested.
   • Subject-led: the user requests or supplies a foreground subject of any kind. When one request asks for both modes, classify and generate each output separately. Do not let the reference choice, palette routing, or contrast logic from one output leak into the other.
3. Select exactly one mode-specific style-only reference for $imagegen:
   • Background-only: use the canonical PNG named by artifact-template.json.
   • Subject-led: use assets/subject-reference.png instead of the canonical background PNG. Do not pass both template references unless the user explicitly requests a comparison or blend. Never use either one as an edit target or composition to trace. When the user supplies a subject image, pass that image separately as the content/edit target and identify its role clearly in the generation brief.
4. Translate the user's request into the relevant rules below. In subject-led mode, treat the requested subject as a conceptual seed: describe only its broad direction, mass balance, placement, and color-temperature role. Explicitly discard anatomy, taxonomy, surface detail, photographic realism, and any assumption that the subject must grow upward from the bottom edge. Route generic color-tone requests according to the subject–field contrast rules instead of tinting the whole image one hue. Do not invent text, logos, UI, objects, or factual content merely to fill the composition.
5. In subject-led mode, resolve placement before calling $imagegen. If the user gives a position, follow it. Otherwise run the placement-card procedure below, write the selected card into the generation brief as a hard COMPOSITION LOCK, and make placement independent of the reference image, subject type, and palette. Do not use words such as “random,” “asymmetric,” or “ample negative space” as a substitute for an explicit card.
6. Visually inspect the result. Unless the user explicitly requests recognizability, reject it if the subject is instantly identifiable as a literal object or species at thumbnail size; if it contains anatomy, surface texture, a sharp focal plane, or macro-photography cues; if subject and field share the same dominant hue or temperature family and merge at thumbnail size; if its subject centroid, entry edge, or empty-space direction violates the selected composition card; if it repeats a recent subject-led layout when eligible alternatives existed; if it copies a reference layout; or if it loses the soft grain and low-frequency color language.
7. Return the final image. If the user asks for variants, make each one compositionally distinct while keeping the same restrained style grammar.