prompt-to-asset: Skill for Claude Code

Brand consistency

Engaged when a user is generating more than one asset for the same brand, or when an existing brand.json / brand.md is in play. Wraps asset_brand_bundle_parse, asset_generate_*, and validation.

BrandBundle shape

palette:
  - { name: primary,   hex: "#0A1F44", usage: mark }
  - { name: accent,    hex: "#FF6B6B", usage: highlight }
  - { name: neutral,   hex: "#FAFAFA", usage: background }
typography:
  primary:   { family: "Geist Sans", fallback: "geometric sans-serif, bold" }
  secondary: { family: "Geist Mono", fallback: "monospace" }
style_refs: ["assets/brand/ref-1.png", "assets/brand/ref-2.png"]
lora:       "assets/brand/brand.safetensors"       # optional Flux/SDXL
sref_code:  "--sref 1234567890"                    # optional Midjourney
style_id:   "uuid-of-recraft-style"                # optional Recraft
do_not:     ["drop shadows", "gradients", "photorealism"]
logo_mark:  "assets/brand/mark.svg"                # canonical mark for composition

Call asset_brand_bundle_parse({ source }) to build this from a brand.md, brand.json, DTCG tokens.json, or AdCP spec.

Palette enforcement per model (strongest → weakest)

Fallback that always works: K-means remap in LAB space + ΔE2000 validation. asset_validate flags drift; post-process recolors nearest palette entries.

Style reference management

• Canonical anchor — one .png or .svg that defines the brand look. Required for consistency across >3 assets.
• Accepted-asset promotion — after an asset passes validation + user acceptance, add it to style_refs[] so subsequent generations see richer brand context.
• CSD similarity score — Contrastive Style Descriptor embedding comparison. Threshold: ≥0.72 pass, 0.60–0.72 review, <0.60 fail. Tier-2 validation.
• Model-specific handles:
  • Midjourney: --sref <image> (loose) vs --cref <image> (character lock) vs --mref <image> (object lock). Use --sw 250 for tight style, --sw 100 for loose brand.
  • Recraft: style_id is a persistent UUID bound to a trained brand style.
  • Flux + IP-Adapter: pass reference image; set ip_adapter_weight: 0.7–0.9 for style, lower for just color.
  • SDXL + LoRA: trigger word in prompt + lora_scale: 0.7–1.0.