jacky-illustration

Paper-Blue Sprout Diagram

Core Positioning

Use a visual language of "warm paper background + Jacky blue + handwritten text + hand-drawn relationships + Sprout" to handle three levels of information density:

Scene metaphor: turn an idea into an image scene that can be understood at a glance.
Knowledge diagram: clearly explain a core piece of knowledge and its composition, mechanism, or comparison.
Information visualization: organize sequence, causality, dependency, branching, cycles, time, or data relationships into a clear reading path.

The three modes are information densities, not three sets of art styles. Aspect ratio is an independent choice and is not bound to the mode.

Default Interaction

The user only needs to provide the content to be expressed; the mode, nodes, page count, reading path, verbatim text, Sprout actions, and blue focal points are all extracted and planned by this Skill, without handing these internal tasks over to the user.

When the user has specified the mode, aspect ratio, page count, or purpose, follow the explicit requirements; automatically decide the unspecified parts.
When the user says "illustrate, generate, make an image, output, see the effect," or directly provides content after invoking this Skill, first explain in one sentence "recommended mode | aspect ratio | estimated number of images," then generate directly without waiting for confirmation.
Only when both modes are reasonable and the choice would significantly change the information trade-offs or the number of finished pieces, ask one brief multiple-choice question; at the same time, give the recommended option and a one-sentence reason. Ask at most once per task.
When the user says "analyze first, only want the plan, do not generate," output only the recommendation and plan, and do not call the generation tool.
Do not ask the user for internal parameters such as "3–5 nodes," "verbatim text list," or "which information to keep"; first extract them yourself from the user's content. Only ask for missing information when the original text lacks key facts, contains conflicting data, or makes it impossible to safely determine the real UI.
For long articles, different paragraphs can be routed separately; the entire piece is not required to use only one mode.

Fixed Reading Before Generation

Before each generation, read:

Then read only one description for the selected mode:

Scene metaphor:references/scene-metaphor.md
Knowledge diagram:references/knowledge-diagram.md
Information visualization:references/information-visualization.md

Do not load the detailed descriptions of all three modes at the same time. After generation, read references/qa-checklist.md。

Step 1: Content Modeling

Answer internally:

After reading, does the reader only need to feel one viewpoint, or need to understand a knowledge structure?
Is the content focus "what it is like," "what it is," or "how it works"?
Which words, numbers, dates, units, and relationships must be preserved verbatim?
Which node, if deleted, would break understanding?
Where should Sprout take action to help understanding rather than serve as decoration?

First write one sentence for the visual task:

这张图要让读者在 3 秒内看见______，随后理解______。

If it cannot be written clearly, continue refining from the original text first; only ask the user when key information that would change the meaning of the image is missing.

Step 2: Route by Information Density

A. Scene metaphor | low density

Express only one viewpoint, emotion, pain point, conflict, or cognitive turning point.
Even if the text is covered, the main meaning can still hold up through objects and actions.
Text is only used to state the theme, not to explain the structure.

Constraints: one core scene, one main action, one main object or compact object group, 0–3 short labels.

B. Knowledge diagram | Medium density

Answer "what it is, what it is made of, what the core mechanism is, and how two states differ."
Need to present 3–5 necessary concepts relatively fully.
Relationships exist, but readers are not required to strictly follow steps.

Constraints: one central proposition, 3–5 semantic nodes, 4–8 short labels, at most one main relationship skeleton.

C. Information visualization | High density

Answer "how it operates, in what order, and how they affect each other."
The core is sequence, causality, dependency, branching, loops, timelines, comparison dimensions, or data relationships.
Titles, steps, dates, quantities, units, or labels must be precise.

Constraints: first extract information modules and automatically plan the number of pages; each page usually carries 4–6 readable modules, one communication task, and a clear reading start and end point. When content is overloaded, split pages; do not shrink text and force it in.

Priority in case of conflict

Precise steps, dates, values, units, branches, or dependencies: information visualization.
No strict order, but need to explain 3–5 components or mechanisms: knowledge diagram.
Only need to establish an intuitive understanding or emotional resonance: scene metaphor.

Do not fix "tutorial" into a certain mode: conceptual explanations in tutorials can use knowledge diagrams, while operation sequences and logical paths go into information visualization.

Step 3: Choose the aspect ratio

All three modes support:

16:9 horizontal: article body, video illustrations, wide processes.
4:3 horizontal: general knowledge cards, more compact diagrams.
3:4 vertical: mobile reading, Xiaohongshu, multi-segment vertical paths.

When specified by the user, strictly comply. When unspecified, choose according to the usage scenario; if still unclear, default to 16:9. Changing the ratio only allows rearranging the composition; it must not delete necessary information or change the mode.

Step 4: Planning

Before generation, internally form:

Mode and basis for selection
Aspect ratio and usage scenario
One-sentence visual task
Information blocks or semantic nodes, arranged in reading order
Yazai's action and position
One main blue focal point and the core information it carries
2–4 secondary blue emphases and the upper limit of blue area
Verbatim text that must appear
Content explicitly not to be drawn

These plans are used internally by default. Only output the plan when the user asks for a plan, without generating; when the user asks for illustrations or directly provides content, give one sentence of recommendation according to "default interaction" and then directly call the built-in image generation tool page by page.

Step 5: Generation

Each image carries only one communication task.
Always use assets/ip/yazai-paper-blue-anchor-v1.png as the first identity and medium reference.
Use the paper-blue master specified by the selected mode file as the second reference, only for art style, composition tendency, and information density.
assets/ip/yazai-canonical-front-v1.png is used only for identity structure verification and identity drift repair, not as a regular generation style reference, to avoid bringing the 3D plastic texture back to the paper surface.
Other Yazai images can only be referenced for action, and must not override the primary identity reference; old infographic cases can only be referenced for structure, and must not override the paper-blue master.
Prompts are arranged as "information structure → main subject of the image → Yazai action → exact text → style and avoid items".
Generate and check one image at a time; do not generate a nine-grid.
List the required text verbatim, character by character, requiring each item to appear only once, with no other text added.
Each image must have exactly one primary blue semantic focal point visible at first glance; blue must carry keywords, core nodes, conclusions, or key actions, and must not appear only in arrows and decorative lines.
The Yazai mask must retain one-sided reflection to express the curved surface: all visible highlights must be concentrated on the same lit side, using one continuous, relatively wide, soft arc-shaped reflection that can naturally fall off on the same side; light spots on the other side, isolated dots, left-right symmetrical bright lines, or figure-eight eyebrows are prohibited. When the character is very small, simplify the reflection shape, but do not remove it entirely.
When generating information visualizations, any incorrect dates, values, units, step directions, or labels must be fixed before delivery.

Generation and cleanup layering

The generation tool is responsible for composition, information structure, character action, and overall material; do not treat it as an eraser for fixing individual pixels or small blemishes.
After the first round of generation, first list all structural, text, identity, reflection, and background issues in one go, then decide the scope of repair. Avoid continuous full-image redraws of "remove a bit, add a bit".
Only structural, informational, or overall style errors permit full-image regeneration; text, local identity, or single-object errors should only edit the corresponding area and lock the rest of the content.
For isolated white dots, small spots, dirty edge pixels, and other tiny defects whose area is clearly smaller than the image, use true local masking, a healing brush, or deterministic pixel cleanup; do not regenerate the entire image for this. If the current tool cannot perform precise local processing, keep the last structurally correct original image and state that pixel-level cleanup is needed, without continuous trial and error.
Perform at most one generative full-image repair for the same image. If, after repair, the remaining issues are already local defects, immediately stop generative repair and switch to local cleanup.
Any generative repair must start from the last original image with correct structure and text, and must not continue accumulating on top of the previous repaired image, to prevent paper texture, shadows, and glyph shapes from drifting round after round.

Step 6: Delivery

Before generation, use only one sentence to state the recommended mode, aspect ratio, estimated number of images, and, if necessary, one sentence about the visual task, without elaborating internal reasoning. After generation, show the images and report:

Image purpose and order
Final file paths
Generated text or data still requiring manual review
The number of generative repairs used, and whether local pixel cleanup was performed

Save within the workspace to:

assets/<topic-slug>-yazai-visuals/

Name in order; do not overwrite old assets unless the user explicitly requests replacement.

Boundaries

Visual covers are outside the scope of this Skill; cover tasks should use a dedicated cover workflow.
Real button positions, menu paths, and product interfaces must use real screenshots; do not let generative models fake the UI.
Do not fabricate facts, dates, citations, sources, brands, logos, metrics, or success states.