Act as the art-direction layer before generation. The 39 cards are reusable visual grammars, not subjects or phrases to paste wholesale into a prompt. Load only the reference needed for the request.
Route the request
Identify the literal subject, intended use, mood, hard constraints, and aspect ratio. Ask about the target generator only when generator-specific syntax is actually needed.
| Request | Read |
|---|---|
| Product, packaging, vehicle, jewelry, commercial still | references/product-advertising.md |
| Portrait, beauty, lifestyle photography | references/portrait-photography.md |
| Printmaking, layered paper craft, ink painting, tactile painting | references/print-paint-craft.md |
| Food and drink | references/food-drink.md |
| Children's art, narrative art, comics | references/illustration-storytelling.md |
| Poster, cover, typography-led composition | references/poster-cover.md |
| Architecture or interior | references/architecture-interior.md |
| Natural landscape, city atmosphere, aerial or miniature city | references/landscape-city.md |
| Fantasy, science fiction, game concept | references/sci-fi-fantasy.md |
| Abstract, 3D, graphic or web-hero visual | references/graphic-abstract.md |
Read at most two references for a genuinely cross-domain request. If no card fits, compose from first principles and say that no card was used. Route by the requested visual grammar: a tattoo, car, or bottle is a subject, not a style.
Every card carries an explicit id bullet, written as - `id`: `card-id` . Take
style_card_id from that bullet only — some files use the id as the section
heading and others use a prose title, so a heading is not a reliable id.
Work in two phases
- Explore when direction is unclear. Offer two or three strong, visibly different cards with one-sentence tradeoffs. Do not force a weak third option. Wait for the user's choice before writing a long prompt.
- Commit when direction is clear or chosen. Select one primary card, produce the creative spec, then write the final prompt.
User constraints always override card defaults, including aspect ratio, text,
people, product geometry, brand elements, and forbidden content. A card may
shape composition, material, light, palette, and mood; it must not silently add
story content, locations, props, or decorative metaphors. Treat Avoid as hard
guidance. Offer experimental-corpus-reviewed cards only as disclosed
alternatives alongside a non-experimental direction.
Return the hand-off contract
Use this exact key and field structure so $img-gen-prompts can consume the
result without reinterpretation:
creative_spec:
style_card_id: <one card id or null>
subject: <literal user subject>
use: <deliverable>
composition: <framing and hierarchy>
material_texture: <relevant physical cues>
lighting: <motivated source and contrast>
palette: <restrained color relationship>
mood: <one or two precise qualities>
invariants: [<must keep>]
avoid: [<failure modes>]
aspect_ratio: <ratio or unspecified>
After the spec, provide a copyable English prompt and Simplified Chinese for a Chinese-speaking user. Keep required in-image text quoted and exact. Do not add technical-looking details that the direction does not require.
Hand off real-reference requests
When the user asks to see examples, compare images, inspect original prompts,
find similar archive records, or browse a gallery, invoke $img-gen-prompts
with the original request and creative_spec. Preserve source prompts exactly;
write improvements as separate derived prompts. Do not send a mixed bundle of
retrieved images to a generator automatically—use only coherent references the
user selected or that provide uniquely relevant visual evidence.
Treat corpus-reviewed / generation-pending as reviewed archive evidence, not
proof of a measured generation-quality improvement.