asset-generation（资产生成）

Asset Generation

Use this skill for logos, icon marks, hero/section imagery, and texture motifs.

Trigger

Run when user asks for:

• logo exploration or logo revisions
• better visual assets in design proposals
• proposal-specific imagery
• provider comparison (Recraft, OpenAI, Gemini/Nano Banana)

Core Rules

1. Use all three providers during exploratory rounds unless user opts out:

- Recraft - OpenAI Image (gpt-image-1) - Nano Banana Pro (Gemini image model)

1. Generate assets in proposal context, not as detached global batch.
2. Exploration width must match design uncertainty:

- Early rounds: high style/color diversity - Late rounds: narrow around survivors

1. Logos must pass favicon/app-icon constraints:

- flat, simple, low-part count - legible at 16px/24px - works one-color and full-color

1. If provider outputs are noisy/illustrative, use a deterministic geometric SVG fallback.
2. Treat user-provided logo references as quality primitives only, never literal templates.
3. Never produce marks that can be confused with existing company logos.

Reference Archetypes (Quality North Star)

Target mark families like:

• modular rounded blocks
• orbiting dot constellation
• constrained grid/mosaic tiles
• minimalist key/keyhole symbol
• 3x3 dot matrix
• hex-aperture emblem

Avoid mascot illustrations, scene composition, and decorative line clutter.

Reference Use Policy (Hard)

When user provides inspiration logos:

1. Extract only abstract qualities:

- simplicity level - geometry type - corner behavior - stroke/fill balance - negative-space strategy

1. Do not reuse exact silhouette, arrangement, or recognizable motif from references.
2. Do not preserve source colorways unless the user explicitly asks.
3. Generate at least 4 structural families per batch before narrowing.
4. Reject anything that looks like "same logo with minor tweaks."

This is a hard requirement. "Inspired by" means transferable principles, not copied form.

Complexity Budget (Hard Limits)

Reject marks that exceed:

• 8 primary primitives
• 2 stroke widths
• 2 brand colors + 1 neutral
• more than 1 internal detail per major shape

Accept only marks that remain recognizable at 24px without anti-aliasing blur.

Phase-Aware Breadth

Phase A: Wide Exploration (no clear winner yet)

Per proposal, generate at least:

• 1 Recraft concept
• 1 OpenAI concept
• 1 Nano Banana concept

Diversity matrix (required across batch):

• Geometry: rounded / angular / mixed
• Weight: monoline / medium / bold
• Symbol strategy: abstract / mic-derived / node-derived
• Palette families: warm, cool, neutral, high-contrast, muted

Phase B: Convergence (2-3 survivors)

Per proposal, generate:

• 2 focused variants (best provider + fallback provider)
• tightly constrained palette and stroke system

Provider Routing

Logo Prompt Contract

Always include:

• no text
• no mockup
• no shadows
• no gradients (unless explicitly testing gradient direction)
• centered symbol
• flat icon

Template: [brand intent], [shape language], minimalist flat icon logo, no text, no mockup, no shadows, no gradients, centered symbol, transparent or plain background

Proposal-Context Workflow

For each proposal 7a..7h:

1. Read proposal DNA + palette + typography mood.
2. Generate 3 logos (Recraft/OpenAI/Nano Banana) using that proposal context.
3. Pick best mark for that proposal only.
4. Embed into that proposal HTML and tune surrounding spacing/contrast.
5. Keep alternates for review.

Do not force one logo family across all proposals during exploratory rounds.

QA Gates (Reject if fails)

• noisy illustration masquerading as logo
• too many micro-details for favicon scale
• style clashes with proposal typography/system
• same colorway/style duplicated across whole batch
• visible gibberish text baked into background assets
• mascot/character or scene illustration instead of symbol mark
• more than one focal object in a logo frame
• obvious resemblance to known logos or user-provided samples
• structural near-duplicate across proposal logos

Deterministic Fallback

When model outputs fail QA, generate a clean SVG mark from geometric primitives. This is not optional for quality: use it to guarantee a production-safe floor.

Output Layout

Use proposal-scoped paths:

• .design-evolution/gen-N/assets/logos/7a/recraft-1.svg
• .design-evolution/gen-N/assets/logos/7a/openai-1.png
• .design-evolution/gen-N/assets/logos/7a/gemini-1.png
• .design-evolution/gen-N/assets/textures/7a/*.png

Chosen production candidate:

• .design-evolution/gen-N/assets/logos/7a/final.(svg|png)

Minimal Command Patterns

Recraft (symbol mark)

python3 "$ENGINE" --repo "$REPO" recraft icon \
  "minimal flat icon mark for Vox Cloud, abstract mic+node, no text, no gradients" \
  --colors "#1E5CFF,#111827" --n 3 --out "$OUT"

OpenAI (gpt-image-1)

curl -sS https://api.openai.com/v1/images/generations \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-image-1","prompt":"minimal flat icon logo mark, no text, no gradients, no shadows","size":"1024x1024","background":"transparent"}'

Nano Banana Pro (Gemini)

curl -sS "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-image:generateContent?key=$GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"contents":[{"parts":[{"text":"minimal flat icon logo mark, no text, no gradients"}]}],"generationConfig":{"responseModalities":["TEXT","IMAGE"]}}'