Cream paper background + a workflow narrative key visual where three textures coexist. The core is not any single material, butthree rendering registers each holding their own boundary within the same image。
Style DNA (seven items, unchanged for any theme)
- Three registers coexist and do not erode each other: flat UI cards (matte off-white, rounded corners, diffuse shadows, thin-line icons, no gradients) / realistic skeuomorphic props (visible fabric weave, metal, paper edges, realistic contact shadows) / semi-realistic painted figures (warm skin tones, soft brush edges, fine canvas grain, not photographic) — the boundaries between the three must be distinct; if any texture spreads into another register, the style collapses;
- The background isa single cream paper color, with only very light paper grain, zero clutter and zero gradients;
- A single purple runs throughall connecting lines, icons, checkmarks, and badges; a highly saturated warm color (orange) is allowed to appear only once in the entire image (sticky note);
- Connecting lines arerounded right-angle purple workflow lines + small solid arrows, a single clear reading order, one feedback loop;
- The core metaphor is carried byone realistic skeuomorphic prop(card catalog box, wooden drawer, type case…), it is the texture anchor of the entire image;
- Figures and the flowchartshare the same physical space, attention falls on the diagram,the main figure does not look at the camera— this is the mechanism that stitches the two worlds together. Number of people, position, pose, action, and cropping method are all free;
- Large serif headline (near-black ink green) + small serif tagline in whicha single word switches to the brand purple。
Item 1 is the easiest to lose: the model will default to unifying the entire image into one texture (all 3D or all illustration). Item 6 is next — once the figure looks at the camera, the image breaks into two halves.
Diversity discipline
The seven DNA items are fixed,the staging must change for every image。
First-round five-image test: the separation of the three registers, the single-color throughline, and the text rendering all held up, but the layout, figure position, pose, cropping, and desktop props of the five images were almost identical — because the initial master template wrote one staging decision from the source image as a fixed value.
Now staging is split into three module libraries (layout / figures / desktop props, see references/from-scratch.md）。Each generation draws one item from each of the three tables, and different from the most recent one; when producing as a group, first list the draw results for the whole group, confirm there are no duplicates, then compile.
Mode discrimination
- The user givesa themeand wants to generate from scratch →
references/from-scratch.md(master template + props/color palette/figure module libraries) - The user wantsa version that can be sent to someone else as a whole block, where the other person only needs to fill in the theme to use it →
ARTICLE-COPY.md - The user gavea reference image in the same styleTo make the same thing → still use the master template; only fill the content decisions from the reference image (what text to write, how many cards, what props) into the empty slots, and use the master template's production methods for everything.
Prompt writing rules
Seven labeled sections, in this order:Use case / Scene / Subject / Key details / Composition / Text in image / Constraints。
Merge exclusions into the Constraints section,do not output a separate negative prompt block— GPT Image 2 has no negative channel.
Put on-screen text verbatim in quotes and give its position and font characteristics, add render all text verbatim, no extra words. Lots of text is an inherent property of this style (title + card labels + badges + tagline); the shorter the copy, the more stable it is, and keep each card label to 1–2 words.
Default aspect ratio is 5:2 landscape (for hero use); 16:9 is usable; square images squeeze the left-right narrative and are not recommended.
Status and known risks
The first round of five images has been tested（2026-07-31）。
Verified as working: separation of the three registers, a single accent running throughout, the four color palettes, and English text rendering (including small card text) all passed on the first try.
| Risk | Status / how to fix |
|---|---|
| Grouped images look exactly the same | Happened in the first round. The scheduling has been split into a module library; each image draws one item from each of the three tables with no repeats |
| Chinese glyphs | Not verified— all five in the first round used English. If you really want to use Chinese, test it according to the item in evals; if it breaks, first reduce the title word count |
| The three registers are unified into a single texture | Did not happen in the first round. If it happens, add material nouns to the register that was assimilated, changing only one place at a time |
| Characters drawn as photos or looking at the camera | Did not happen in the first round. If it still happens, move the character action sentence to the start of the Subject section |
| Text rendering errors | Did not happen in the first round. If it happens, first reduce the total amount of copy, then shorten the labels card by card |
| Flow lines piercing cards, arrows pointing randomly | Did not happen in the first round. If it happens, reduce the number of nodes (≤6) and add no embellishment |
Guardrails
- Describe the style using visual characteristics; do not name artists and do not reproduce real brand logos (icons such as Discord/Slack are all generalized into generic shapes such as chat bubble / envelope).
- The characters are fictional generic figures; if the user asks to draw a real person, that person's consent is required.
- For image deviations, first suspect that the Subject slot is not specific enough; do not touch the DNA fixed block; change only one place per round.