Compilea statementintoaprompt that can be handed directly to an image model. Style is locked: a white void overlaid with a very faint one-point perspective grid as the substrate, with a realistic subject placed inside, the subject itself bringing out pure black as the frame's only black anchor, and an oversized serif main headline pressed across the top of the image.
Core principle:One sentence, one object, one block of black. The premium feel of this style does not come from having many elements, it comes from tonal purity and restraint — white is paper white, black is ink black, and only one thing in the frame is lit. Add a second subject, let gray tones crawl all over the frame, and the style collapses immediately.
Perspective grid substrate
The background is not a blank expanse, it is aone-point perspective space drawn with hairline light gray lines: the floor, ceiling, and both side walls are all covered with a grid, all lines converging into a vanishing point at the exact center of the frame, and the cells get larger the closer they are to the edge of the image. It props up depth out of pure white, giving the subject somewhere to stand.
Three boundaries; cross them and it turns from substrate into a second subject:
- It can only be the lightest tier of gray, hairline thickness, never thickened, never darkened, never smudged into a surface;
- It does not cross the headline, the grid breaks in the upper horizontal band where the headline sits, leaving a clear corridor;
- It stays behind the subject, the subject presses in front of the grid and does not compete with it for attention.
When the subject needs a different floor (dot matrix, water surface, pure white), change only the floor item; the grid on the walls and ceiling does not move. When the subject fills the frame (for example the underside-of-a-bridge type), the grid is visible only in the remaining white void.
Aspect ratio
5:2 ultra-wide horizontal format, locked. Use 3200×1280、quality highfor the generation parameters — hairline serifs and the light gray perspective grid both eat resolution, and at low resolution the grid smudges into a layer of gray fog, and the tonal purity of the whole set collapses with it.
Other aspect ratios are unverified; change only if the user explicitly requests it, and remind them that the white-space ratio needs to be re-measured.
Two language registers
Choose one at a time, and keep the whole set consistent. The two registers areequal in status, Chinese is not a downgraded version of English.All rendered text is serif, this is the identity characteristic of this style.
- EN: main headline in a Didone serif (Bodoni type), extreme thick-thin contrast + hairline serifs; descriptor in small-size wide-tracking serif uppercase.
- CN: main headline in a high-contrast Chinese serif,specifically Source Han Serif, thin horizontals and thick verticals, sharp triangular serifs; descriptor in small-size wide-tracking same family.
Whichever language the user's headline is in, use that register; do not translate and do not polish.Didone Write it only for Latin— it is a Latin typeface lineage name, Chinese has no corresponding bloodline, and writing it on Chinese is the same as not writing it. For the Chinese character-count ceiling and verification method, see compile-contract.md Chinese rendering special.
Read
- Prompt compilation: read
references/compile-contract.md(iron rules + locked kernel + two language registers + fixed skeleton). - Picking a main installation, wanting finished reference, or batch-making as a group: read
references/examples.md(main subject vocabulary + verified finished-product prompts).
Input
Infer from user input, no form-filling required:
- Main title: copy the user's original text verbatim. EN ≤4 words (6 max), CN ≤8 characters. This is the first visual focus of the image.
- descriptor: a small text label above the main title. EN ≤3 words, CN ≤6 characters. If the user didn't provide one, make one up based on the main title's context.
- Main installation: compress this judgment into a concrete object that can be named. Consult
examples.md's main subject vocabulary. - Register: follow the main title's language.
Workflow
- Read the judgment: first state clearly whatrelationshipthis sentence describes (opposition / load-bearing / accumulation / chain reaction / choice / compounding...); the relationship determines the installation.
- Choose the installation: consult the main subject vocabulary. One image can have only one dominant installation; it may carry internal layers (one main object + several supporting players), but cannot have two co-equal subjects.
- Set the black anchor: in this installation,which part is pure black? If there is no clear black anchor, change the installation—this is this style's load-bearing wall.
- Set the illuminated object: only one thing in the entire image is lit; everything else sinks into white or black. This is the second load-bearing wall.
- Compile the prompt: output the finished product according to
compile-contract.md's fixed skeleton, no placeholders, no menu, zero Chinese in the body (except the rendered title). - Output: register + one complete prompt + (optional) one sentence on the reason for the installation choice. Omit the explanation when the user only wants the prompt.
When producing as a group: the style base (the three sections Key details / Text specs / Constraints) must be written verbatim intoevery singleprompt—the model only sees a single prompt, and shared rules are not automatically inherited. Only the Scene plot, Subject, and text change.
Quality Gates
- Tonal purity: the background is paper white, the anchor is ink black; gray may appear only in the narrow transition on the subject's own surface and on the base grid, and must never become an area's tone. If the output looks gray, it fails.
- The grid is the lightest tier: the perspective grid must be noticeably lighter than any gray on the subject. Thickening, darkening, or blurring the grid into a surface, or lines crossing the title, all count as failures.
- Single illumination: within one installation there is only one lit spot; the rest sinks into black. Lighting multiple spots will chop the black area into a gray colonnade—this is the most common collapse point of this style.
- Whitespace ≥50%: the subject occupies the lower half, the title occupies the upper 30%, and the middle is left clear. The subject must not press on or block the text.
- Flat text: rendered text is a flat solid-black graphic, with no shadow, no dimensionality, no perspective. A large title with shadows instantly becomes cheap.
- Text precision: only the two specified lines appear, and a verbatim clause is mandatory. If a third line of text, a watermark, or a logo appears in the image, it fails.
- Chinese character-by-character verification: After the CN register image is produced, check the glyphs. If there are typos or fabricated characters, redraw them exactly as they are; if there are two extra images, pick one. Do not preset garbled text, and do not fall back to leaving blank space for later typesetting.
- This skill only produces prompts and does not fabricate generated image results.