diagram

name	diagram
description	Generate professional academic structure diagrams, framework charts, and logic architecture visualizations. Default output is direct gpt image-2 image generation for both previews and final bitmap diagrams. Use HTML only when the user explicitly asks for an editable source file; do not generate HTML as an intermediate step for PNG images.

Generate professional structure diagrams and framework charts for academic papers, technical reports, and presentations. Use a formal Chinese academic visual language to show relationships between concepts, technologies, processes, resources, and outcomes.

Output Format

Default and final bitmap output: Generate images directly with gpt image-2. Use this for previews, drafts, selected styles, and final PNG/JPG-style deliverables.

Do not use HTML as a PNG pipeline: Do not create HTML first and then export PNG unless the user explicitly asks for an HTML source file. For ordinary image requests, generate the image directly with gpt image-2.

Editable source output: Only when the user explicitly asks for editable source, HTML, manually adjustable source, or a file that can be edited later, generate one self-contained HTML file under the output directory. HTML is an optional editable source deliverable, not the default image-generation path.

Only generate HTML when explicitly requested. Other editable document or vector source formats are not required outputs.

Image-2 Output Rules

Use gpt image-2 for direct diagram images.

Image generation requirements:

• Generate the complete visual diagram directly as a bitmap image.
• Prefer 16:9 wide composition unless the user requests another ratio.
• Use blue academic/reporting style by default.
• Use clear hierarchy, strong title area, balanced modules, elegant arrows, smooth curves, professional card frames, and polished light/shadow.
• Prefer abstract diagram structures over decorative literal objects: cards, bands, rings, layers, funnels, paths, timelines, matrices, pillars, and arrows should carry the logic.
• Keep icons small and secondary. Use icons as module markers only, not as the main visual body.
• Avoid large literal icons or symbolic mascots such as giant shields, trophies, medals, target boards, rockets, crowns, hands, cartoon objects, oversized buildings, or decorative badges unless the user explicitly requests them.
• Use Microsoft YaHei / ???? style direction when prompting, but accept that generated text may not be perfectly editable.
• When Chinese text accuracy is critical, keep diagram text concise and provide exact labels in the prompt.
• Avoid watermarks, brand marks, UI chrome, fake app screenshots, messy backgrounds, or random English filler.
• Do not generate Nano banana, Gemini Nano banana, logos, or corner marks.

HTML Output Rules

Use HTML only when the user explicitly asks for editable source, editable diagram, source file, HTML, vector-like editing, or a file that can be manually adjusted later.

HTML requirements:

• Generate a single standalone .html file in the output directory.
• Put CSS inside the file; do not rely on external CSS files.
• Use font-family: "Microsoft YaHei", "????", Arial, sans-serif for all Chinese text.
• Use HTML/CSS layout primitives for structure; embedded vector markup may be used inside the HTML for complex arrows, rings, funnels, and icons.
• Avoid external assets, external fonts, remote scripts, remote CSS, tracking code, or web dependencies.
• Do not automatically export HTML to PNG unless the user explicitly requests both HTML and PNG from the HTML source.

Visual Clarity Rules

For Direct Image Generation

• Use gpt image-2 for visual quality, layout sophistication, arrows, smooth curves, glow, card frames, and overall composition.
• Keep text concise to reduce image-generation text errors.
• If the user provides long source text, first compress it into diagram labels and short explanatory phrases.
• If exact Chinese text matters more than style, tell the user that direct image generation may still require manual correction or a separate editable HTML source.

Text Accuracy

• For direct images, provide exact Chinese labels in the image prompt.
• Prefer short labels, short titles, and short bullet phrases.
• Avoid dense paragraphs inside generated images.
• For numeric indicators, keep each number close to its label and use red only for key numbers.

Icons, Arrows, and Frames

• Let image-2 handle polished smooth arrows, curved funnels, rings, card frames, and light/shadow in direct image output.
• Do not make large icons the visual focus. Default to small line icons inside circles or card headers.
• Avoid giant shields, trophies, medals, target icons, decorative emblems, or oversized literal objects because they often look strange and weaken the logic diagram.
• For central concepts, prefer clean text cores, abstract circles, rounded rectangles, layered platforms, or simple geometric hubs instead of a large pictorial symbol.
• Use 0-1 icon per module when useful; if the diagram already has many modules, omit icons and use numbered labels instead.
• Do not use low-resolution screenshot icons or raster frame decorations as source material.
• If using HTML by explicit request, use crisp CSS/vector icons and frames.

Workflow

1. Understand the User's Content

Before generating any diagram, understand:

• Topic/Theme: The main subject of the diagram
• Subtitle: A descriptive subtitle explaining the diagram's purpose
• Core Concepts: The central ideas or technologies
• Categories: Major groupings or dimensions
• Sub-items: Detailed items under each category
• Relationships: How concepts connect - flow, hierarchy, cycle, support, conversion, feedback, or output
• Key highlights: Which elements need visual emphasis

If the user provides a topic without structured data, organize the content into a concise structure before generating.

If the user provides a reference image, identify what must be preserved: layout, icon style, frame style, color, hierarchy, curved shapes, arrows, and spacing. Use it as visual guidance for image-2 unless the user explicitly requests editable HTML recreation.

2. Choose a Layout Pattern

Consult reference/layout-patterns.md to select the most appropriate layout. Choose the layout that represents the logical relationship, not only the one that looks attractive.

Relationship Type	Recommended Layout	When to Use
Central concept radiating sub-topics	Radial Hub	One core idea with multiple dimensions
Sequential stages or pipeline	Pipeline Flow	Step-by-step processes, value chains
Two complementary forces	Dual-Cycle	Balance, tension, or feedback loops
Multiple parallel categories	Matrix Grid	Capability maps, comparisons, taxonomies
Core expanding outward	Core Expansion	Central principle with derived applications
Bottom-up or layered structure	Hierarchical Layers	Theory to practice, base to outcome
Technical component relationships	Tech Architecture	System design, technology stacks
Closed-loop ecosystem	Ring Ecosystem	Integration, ecology, full lifecycle
Bottom-up strategic support	Pyramid Hierarchy	Goals supported by foundations
Driver or conversion engine	Core Engine	Inputs feeding a central mechanism
Evidence feeding mechanism	Evidence-to-Mechanism Canvas	Foundation/evidence to core logic
Three forces converging	Triangle Convergence	Three supports to one target
Iterative improvement loop	Central Loop Engine	Build, apply, evaluate, improve
Cross-domain collaboration	Upstream-Downstream Synergy	School-enterprise, industry-education linkage
Two-sided comparison	Mirrored Comparison with Bridge	Current base vs. target capability
Many inputs refined into outputs	Layered Funnel Roadmap	Resource construction, standards development

3. Apply the Visual System

Consult reference/visual-system.md for design details. Default style:

• Blue academic/reporting style
• Dark blue headers, bright blue accents, light blue surfaces
• Red only for important numbers or key metrics
• Clear hierarchy: title > section headers > item labels > descriptions
• Microsoft YaHei / ???? style direction for Chinese text

4. Generate the Diagram

Prompt construction algorithm for image output:

1. Compress the source into a title, one core message, 3-5 main modules, and a short output/result band.
2. Choose one layout pattern and name it explicitly in the prompt.
3. Describe the logical shapes first: modules, cards, arrows, layers, rings, paths, or funnels.
4. Add a negative icon rule: no giant shield, trophy, medal, target, rocket, crown, cartoon object, oversized building, or decorative badge unless explicitly requested.
5. Add a readability rule: large Chinese labels, few words, no dense paragraphs, no random English filler.
6. Use small line icons only if they support scanning; otherwise use numbered circles or clean labels.

Construction principles:

• Use one strong main title area.
• Use 2-4 major sections arranged by the chosen layout.
• Use 3-6 sub-items per section when possible.
• Show visual flow clearly with arrows, funnels, rings, layers, or grouping boxes.
• Do not produce flat list posters. When the source is enumerative, compress items into logical forces and show support, conversion, feedback, or output.
• For image output, write a precise gpt image-2 prompt describing layout, style, hierarchy, exact labels, and forbidden marks.
• For explicitly requested editable source, create a single HTML file and make all text, icons, frames, arrows, and diagram structures editable through HTML/CSS.
• When matching a reference image, preserve the visual grammar: icon type, card proportions, curved funnel shape, arrow style, title bar, effect column, and bottom result band.

5. Polish and Verify

After generating, verify:

• Readability: Text can be read at normal viewing size.
• Hierarchy: The core concept is visually dominant.
• Flow: The reading order is obvious.
• Balance: The layout is not visually lopsided.
• Completeness: All user-provided concepts and key numbers are represented.
• Consistency: Similar elements use consistent styling.
• Output rule compliance: Default/final image output uses gpt image-2 directly; HTML is generated only when explicitly requested.
• Visual clarity: No watermarks, fake logos, UI chrome, blurry screenshot elements, or unwanted corner marks.

Layout Templates

Use the matching HTML template from templates/ only when the user explicitly requests HTML/editable source:

• templates/radial-hub.html
• templates/pipeline-flow.html
• templates/dual-cycle.html
• templates/matrix-grid.html
• templates/core-expansion.html
• templates/hierarchical-layers.html
• templates/tech-architecture.html
• templates/ring-ecosystem.html
• templates/pyramid-hierarchy.html
• templates/core-engine.html
• templates/evidence-to-mechanism.html
• templates/triangle-convergence.html
• templates/central-loop-engine.html
• templates/upstream-downstream.html
• templates/mirrored-comparison.html
• templates/layered-funnel.html

When using a template:

1. Read the appropriate template.
2. Replace sample content with the user's content.
3. Adjust sections/items as needed.
4. Save the final editable source as a .html file under output.

Customization Options

Support user requests for:

• Color theme: blue default; red, green, purple, or gold variants when requested
• Logo/branding: only if the user supplies it or explicitly requests it
• Aspect ratio: wide 16:9 default; 4:3 or tall layouts when requested
• Density: compact or spacious
• Style variant: formal academic, modern tech, soft business
• Output mode: direct image-2 image by default; HTML only when explicitly requested

Anti-Patterns

Never do these:

• Generate HTML first just to produce a PNG, unless the user explicitly asks for HTML and PNG from the HTML source.
• Overcrowd the canvas with tiny text.
• Use random English mixed into Chinese diagrams.
• Mix layout patterns without a clear reason.
• Use decorative elements that do not communicate meaning.
• Generate placeholder content.
• Use giant decorative icons as the main graphic, including shields, trophies, medals, target boards, rockets, crowns, hands, cartoon objects, oversized buildings, and ornate badges, unless the user explicitly asks for that symbol.
• Let icons replace the logical structure; the structure must be expressed by modules, arrows, layers, paths, rings, or matrices.
• Use external fonts, remote images, remote scripts, or remote CSS in editable HTML output.
• Use raw screenshots as final visual panels when HTML is requested.
• Use low-resolution raster icons or blurry frame decorations.
• Require any editable source format other than HTML.
• Repeat the same card grid for multi-image deliverables.

Remember: Default image deliverables should prioritize visual quality through gpt image-2. Use HTML only for explicit editable-source needs.