Compilea judgment about a certain platform/productintoaprompt that can be handed directly to an image model. Style locked: near-white studio void, on the right hangs a thick rounded 3D App icon carrying the real brand logo, on the left three layers of Chinese text pressed on the same horizontal axis.
Core principle:One icon, one knockout block, one horizontal axis. This style's cleanliness does not come from having few elements, it comes fromstructural alignment—the text block and the icon share one optical center line, the accent color appears in only one place (the knockout block), and props do not exceed two. Add a second subject, let the accent color scatter to three or four places, let the shadow creep into the text area, and the style collapses immediately.
Aspect ratio
5:2 ultra-wide horizontal board locked. For the image parameters use 2560×1024、quality high——dense small Chinese text and the narrow highlights on the icon's bevel both need resolution.
Other aspect ratios are unverified; only change it if the user explicitly asks, and remind them that the whitespace ratio and the upper limit for text block width need to be re-measured.
Logo protocol (the crux of this skill)
The entire recognizability of this style rests on that one logo, so it has its own four rules:
- Research the current version before writing. Brands change their marks (Twitter → X), multiple versions exist simultaneously (App icon version / horizontal wordmark / dark-background light-background versions), and they may also have just been redesigned. Confirm which version is currently in use, andname the version：
the current X app icon、the current 小红书 app icon。 - Name it directly, do not decompose the glyph. Writing
the X logois enough, the model recognizes it. Break it into strokes, intersections, how the stroke ends are cut—the more detailed you write, the more crooked it draws. - Only supplement the presentation: what color, attached to which face, what proportion of that face it occupies, whether it rotates with the subject's perspective, whether it has embossing or outline.
- Block retired old marks by name into Constraints. The model's default often stops at the old version; X will produce a blue bird, Weibo will produce the old eye mark. This is a required item, not insurance.
There is another rendering discipline written into Constraints:logo must lie flat on the front face, must not wrap with the bevel or stretch and deform—this is the most common bad outcome for labels on 3D faces.
Color register: follow the brand
Colors are not preset,determined by the brand's primary color, but the architecture is fixed as four roles:
| Role | Value | Area |
|---|---|---|
| Base | Near-white studio (cool white or warm white, according to brand tone) | ≈85% |
| Subject three tiers | Front / side / bottom three levels of brightness of the brand primary color,must be given separate values | Icon |
| Accent | Pure black or brand primary color, appearing only in three places in the text | Minimal |
| Secondary text | Cool mid-gray #6B7280 |
Small |
The three-tier value scale is load-bearing. If you only give one brand color, the three-dimensional icon collapses into a flat swatch of color—the front / side / bottom faces must each differ by one tier. When the brand color is a black-and-white system (X, Apple), this matters even more, or the whole icon smears into a single blob.
Inverted block: emphasis device
Emphasized words in a large headlinedon't change color, use an inverted block—a solid rounded rectangle in pure black (or the primary brand color) + white text. This is what sets this style apart from "big text plus colored keywords," and it's also why it still reads the emphasis at thumbnail size.
When there's a block, don't use parentheses.「共用实名」新政策 should compile into the form [创作者分成]新规则 —the block takes on the function of quotation marks.
Read
- Prompt compilation: read
references/compile-contract.md(fixed skeleton + all values + full text of the seven slots). - To choose props, get the brand register, or get finished references: read
references/examples.md(brand register table + prop mapping table + verified finished prompts).
Input
Infer from user input; filling out a form is not required:
- Brand: which platform/product this is about. First research its current logo version.
- Headline: emphasized word + tail word, rhythm is "four to five characters + three characters." The emphasized word goes into the inverted block. Over five characters, the inverted block grows long enough to eat into the right margin.
- kicker: a small label above the headline, ≤4 characters. In a series, it is a constant (e.g. "Platform Watch").
- Subheadline: ≤10 characters. In a series, this is also often a constant.
- Props: choose 1–2 items from the prop table by theme, two max.
Workflow
- Research the logo: confirm the mark version the brand currently uses, and whether there are retired old marks that need blocking.
- Set the three-tier value scale: take the primary brand color and give three color values for front / side / bottom.
- Choose props: check
examples.md's prop mapping table and choose 1–2 items by article theme. Props are a meaning layer, not decoration. - Compress the headline: compress the judgment into "emphasized word + tail word," with the emphasized word in the inverted block.
- Compile the prompt: output the finished piece per
compile-contract.md's fixed skeleton, no placeholders, no menus, zero Chinese in the body text (except rendered text). - Output: one complete prompt + parameters + (optional) one sentence of rationale for prop selection. Omit the explanation when the user only wants the prompt.
When producing as a group: the four sections Key details, Composition, Text specs, Constraints must be written verbatim intoevery singleprompt—the model can only see one prompt at a time, and shared rules are not automatically inherited. Only the brand name, the three color values, the props, and the text change.
Quality Gates
- Logo fidelity: after generating, check whether the logo is the current version, whether it has been deformed by beveling/stretching, and whether it has reverted to an old mark. If any of these three is wrong, the whole image is void and must be regenerated—the entire persuasive force of this style rests here.
- Three-tier value separation: whether the icon's front / side / bottom faces read as three levels. If it collapses into one flat plane, it fails.
- One horizontal axis: The midline of the main heading and the optical center of the icon must be at the same height. A slight deviation and it falls apart—this is the spot most easily lost to the model's free interpretation.
- Shadow does not enter the text area: The key light is at the upper left, the shadow falls to the lower right, and it must never crawl onto the text block on the left.
- Only one emphasis: The reversed-out block is the only accent in the image. The subtitle's
&and the thin line above are secondary embellishments and do not count as a third emphasis. - Props ≤ 2: If there are more than two, the protagonist status of the main subject collapses, and the image degrades from a "scene" to a "pile-up".
- Exact Chinese: Only the specified three lines appear, and the verbatim clause is mandatory. If a fourth line of text, a watermark, or extra punctuation appears in the image, it fails.
- This skill only produces prompts and does not fabricate generated image results.