Compile the content into a warm, tactile object-metaphor still-life image: cut-paper / handmade miniature texture,one physical object stands in for one judgment。
This skill covers both article covers (5:2) and chapter / section illustrations (16:9). Determine which one the user wants, or do both.
Intent detection
Only ask when it is truly impossible to tell; otherwise infer directly from the request:
- "cover" / "cover image" / "main image" / "hero image" → Cover mode, one 5:2 image.
- "chapter images" / "an image for each section" / "chapter images" / gave an article path and said "every section" → Chapter mode, one 16:9 image per main
##chapter. - Only said "illustration", or "both cover and chapters" → do both: first one cover, then one per chapter.
If the user explicitly specified an aspect ratio (4:5, 1:1, 16:9 cover, etc.), use the user's; do not push the default value.
Language confirmation (cover mode only)
Infer the title language from the article or topic itself — a Chinese title gets a Chinese title and subtitle. State in one sentence which one was chosen so the user can change it,do not stop and wait for an answer; only ask when the source truly gives no language signal at all. Any strings printed on props are always English. Chapter images never carry a large title.
Palette
The palette is arole system, not tied to any specific brand. Six roles:
| Role | Function |
|---|---|
| Ground | Main background / main supporting surface, the absolute dominant color of the image |
| Signal | The only main action, key mark, the one state that is lit up |
| Support | Soft supporting cards, panels, light-colored surfaces |
| Accent A | Very small areas of reminder / community accents |
| Accent B | Status dots, time capsules, progress bars |
| Anchor | The only dark physical object that anchors the scene, the structural anchor, and can also serve as the title color |
Color discipline: Ground is absolutely dominant;only one physical object in the entire image is Signal color— a second element that wants it switches to Support; Anchor lands only on one anchoring physical object plus a few small label strips, never spread into a large color block; Accent A / B are only small-area accents; do not introduce extra neon or cyberpunk color schemes.
Materials are completely open— cut paper and cardstock, string and thread, brass and other metals, wood, glass, stone; use whatever the object actually is, any size, and it can be the protagonist. The palette governs color, not material.
The color of every element must be named in the prompt: paper and board are always Ground color unless another role color is explicitly given. If not written explicitly, the model will drift to kraft paper, ochre, or cardboard brown — off-palette and looking dirty.
If the user provided brand colors, map them to these six roles; if not, use the presets in references/palette-contract.md . The mapping rules and presets are both in that file.
Input Reading
If a local article path is given, read the full text first (read explicitly as UTF-8 in case the terminal swallows Chinese). Extract: title, core argument, the strongest point of tension, the promise, the metaphor for each chapter (chapter mode), and any concrete signals, numbers, timestamps, and short labels that can be used as small details.
If no path is given, use the body text or topic provided in the chat.
Output Conventions
If a local article path is given or the user asks for a file, write a standalone Markdown file; otherwise use the same structure directly in the reply.
Cover filename:[文章名]-diorama-cover-prompt.md
Chapter filename:[文章名]-diorama-illustration-prompts.md
By default, save next to the source article unless otherwise specified.
Cover file structure:
# Diorama Cover: [文章标题]
## Source
## Cover Strategy
[选定的物件域、隐喻、焦点实物、标题语言、调色板映射]
## Cover Prompt
[一条完整可直接复制的提示词，标题文字直接写进去]
## Cover Copy Fallback
[只有反复生成都糊字时才需要：留白构图说明 + 后期合成用的准确标题副标题]
## Acceptance Checklist
Chapter file structure:
# Diorama Illustrations: [文章标题]
## Chapter Prompt Map
| 章节 | 物件域 | 焦点实物 |
## Chapter 1 Prompt
## Chapter 2 Prompt
...
## Acceptance Checklist
When outputting only in chat, keep the same structure.
Fixed Rules
- Cover 5:2 landscape, chapter image 16:9 landscape. The target model is GPT Image 2, butthe model name does not go into the prompt body, only stays at the document level.
- Style: paper-cut / handmade miniature still life, light skeuomorphism, high-end editorial design feel—not photography
- Background: Ground color
- Cover text:the main title and subtitle are always paired, never the main title alone. The main title uses a classical serif (Chinese Ming typeface, with a matching Latin serif); the subtitle is one size smaller, sans-serif, placed tightly below it, also in Anchor color. Both are burned into the image. Text strings printed on props stay in English. Chapter images are English only, with no main title.
- Logos, platform icons, product glyphs, and UI controls are all allowed—use them when the content needs them to carry information.
- Solid color, no gradients
- Natural, soft shadows; materials must have a realistic tactile feel. Desk lamp light, directional light, rim light, and glow are all allowed.
Object Selection
Grow the focal object and supporting objects from this content's own metaphors first.Don't reach for the same batch of props every time—pieces about speed and pieces about trust, memory, evidence, or scale should not land on the same object.
In chapter mode, no two chapters in the same article share a focal object, and none reuse one already used on the cover. In cover mode, avoid objects used in the last 2-3 covers of this project.
Define the object domain first, then the specific object, and rotate object domains—don't let all the images in a project, or all the chapters in an article, sink into the same domain. For the full domain library and rotation rules, see references/object-domains.md。
Objects must be real things you can name. Readers recognize the object first, then decode the meaning; an abstract geometric shape wrapped in a material skin doesn't count as an object, and if it can't be recognized, everything after that breaks down. Likewise, abstract quantities (data, performance, better/worse) must land asreadable words and numbers, or as the physical properties of the object itself (height, thickness, size)—miniature bar charts and small icons degrade into texture at thumbnail size.
Only use a laptop or phone when the content truly needs a screen; it is not the default protagonist.
Scene and Effects Discipline
No set building.Do notbuild out landscapes, terrain, or stages (paths, gravel, rocks, vegetation, background sets). A plain Ground base or one with very slight texture is enough.
No set building does not mean fewer props. Prop density is another matter—a dozen objects on a plain ground is still this style. The image is completed through two layers:meaning layer(the hero physical objects, the relationships between them, the annotations printed on them) and theatmosphere layer(desktop accompaniments that carry no semantics at all and only make the surface look real—desk lamp, pen, paperclip, sticky note). Both layers must be present; cut the atmosphere layer entirely and the image goes dry.
Richness comes from variety of kinds, not from copying the same shape many times over—the same shape repeated in an arrangement reads as texture, not as a scene. Hero physical objects need internal layering; a pile of equal-status objects of similar size reads as a pattern, not as an image.
Relationships must be drawn, not implied by juxtaposition—flow direction, routing, sequence, and belonging all need a visible connector (rope, string, arrow, pushpin, continuous arrangement). Objects placed side by side read as unrelated to each other.
Everything stays still: no motion blur, no impact sparks, no splashes, no collision or friction effects, no raised dust.Static light is fine—desk lamp glow, light beams, rim light, glowing markers are all allowed. Objects suggest state through position, patina, and light, never by implying motion.
Text rules
Cover: both the main title and the subtitle go into the prompt, andare actually generated in the image—not just the main title, and don't leave a blank area by default. The subtitle carries the second half of the claim, not a repeat of the main title: the main title names the thing or poses the question, the subtitle lands that judgment (小模型什么时候就够用 / 匹配任务，不是越大越好). For a Chinese title, put the exact Chinese into the prompt, and after generating check for smeared characters; regenerate first; only fall back to a blank composition plus post-production compositing if repeated regeneration still smears.
The row of small capsules below the subtitle isa product banner device—use it only when there are genuinely quantifiable proof points (response time, count, status). In a concept-explainer cover the image already shows the facts, so don't add it. If you really do add it, keep it in English regardless of the title's language.
Chapters: English only, no main title, no Chinese.Text printed on props is the primary carrier of meaning, not decoration—small capsules (4:12、300+、T+24h、Ready), index labels, field names, status lines, engraved nameplates, handwritten sticky notes, and real numbers can all be used. Density is set per piece: text is there to make relationships and data readable, not to fill space.
Every string in the image has to pass this test: does the thing it says already get shown by the objects? Labels that name a role, status, category, or real number hold up; capsules that repeat what the image already makes clear are redundant—delete them.
Prompt writing order
Write each prompt in seven sections, with section labels, in this order. The model name does not appear in the body text.
- Use case: what deliverable this is (cover / chapter image) and the aspect ratio
- Scene: the Ground base, the paper-cut miniature register, the direction of light
- Subject: the hero physical object, and the one relationship that connects it to everything else
- Key details: color-role placement, supporting props, atmosphere layer, materials and shadows
- Composition: positions, text fields, whitespace, framing
- Text in image: the exact quoted string + font specs + position, with
render all text verbatim, no extra words - Constraints: exclusions, merge directly into the body
Every prompt must be complete and directly copyable—don't split key constraints into vague notes for the user to assemble themselves.
Cover skeleton (already tested)
Run multiple covers through in one go using this shape. Reuse the fixed blocks verbatim; fill the square brackets per piece.
Use case: editorial article cover, wide 5:2 horizontal banner, headline and subtitle baked into the image.
Scene: a warm cut-paper craft-diorama still life staged on a plain [Ground 色] desk surface, light skeuomorphic, premium editorial register; no scenery or terrain; [光源与方向] casting short consistent shadows [投向反侧]; everything at rest.
Subject: [主角实物、它的状态，以及把它与其余部分连起来的可见连接物].
Key details: [每件物件是什么材质、承载什么]. [配色块：把纸钉死在 Ground 色，点名唯一那件 Signal 色物件并声明画面里没有别的东西是这个色，安排 Anchor 与 Support 的落点]. Ambient desk props sit in the lower left carrying no meaning: [从 — 回形针、斜放的 Anchor 色笔、带手写记号的 Support 色便签 — 里挑两三件]. Solid flat colours only, [Ground 色] dominant, visible material thickness and clean cut edges, tactile grain, soft shadows.
Composition: wide 5:2 horizontal frame, eye-level still-life view slightly above the desk so [什么必须保持可读]; [主角在右侧三分之二的位置安排]; the left third holds a two-tier text column — headline, then subtitle — in its upper half, ambient props beneath it; keep the frame comfortably filled, no large dead areas.
Text in image: headline "[大标题]" in a classic serif typeface, Ming/Song-style characters[ and a matching serif for the Latin words], [Anchor 色], large, upper left, two lines allowed. Directly below it the subtitle "[副标题]" in a smaller Chinese sans (思源黑体 / Source Han Sans), same [Anchor 色]. [每一串印在道具上的字，逐一写出并指明落在哪件物件上]. Render all text verbatim, no extra words.
Constraints: no people, hands, or faces; no landscape or terrain staging; no gradients; no motion streaks, impact sparks, splashes, or friction effects; no watermark.
Default constraints
Merge into the Constraints section of every prompt,never output as a separate negative prompt block—GPT Image 2 has no negative channel; a constraint in its own block will only be lost or drawn as a positive description:
no people, hands, or faces; no landscape or terrain staging; no gradients; no motion streaks, impact sparks, splashes, or friction effects; no watermark
Acceptance checklist
Check item by item before finishing:
- The cover states
5:2 horizontal, the chapter image states16:9(or the aspect ratio specified by the user); the model name is not in the body - Seven labeled sections, in the correct order; constraints merged into the body, not a separate block
- Cover: main titleandsubtitle both present, a large serif main title + a one-size-smaller sans-serif subtitle beneath it, never only the main title
- Chapter: English only, no main title; every string of text printed on a prop is spelled out verbatim, with
render all text verbatim, no extra words - The paper color is pinned to the Ground color; exactly one object in the whole image is the Signal color; Anchor is not laid out as a large color block
- Every string of text in the image says something the object doesn't; no capsules repeating information already in the image
- The relationships between objects are drawn with visible connectors, not left for the reader to guess
- Both the meaning layer and the atmosphere layer are present
- 1-2 focal physical objects, from an object domain that fits the content, not repeated between chapters or with the cover
- No landscape construction; the image is static—light and glow are fine, motion is not
- All six palette roles are specified with where they fall