doubao-visualization

Unified Visualization

Overview

Treat visualization as a combination of "task goal × material strategy × presentation method," rather than mechanically choosing one of four old modes. First determine what the user wants to explain clearly, then determine how the user's original images, data, and materials should be used, and finally choose one or more complementary visualization modules.

Core principles:Use deterministic rendering for precise information, preserve original image evidence as-is, generated images are only responsible for explanation, and use interaction only when the change itself is valuable.

Use cases

Use this Skill when the user explicitly requests drawing, diagrams, illustrations, charts, annotation, animation, or interaction, or when the task does not explicitly say "visualization" but contains precise data relationships, visual evidence from original images, dynamic changes that need to be observed, complex knowledge structures, or other content clearly suitable for graphical representation.

Do not use this Skill when plain text is clearly clearer, the user explicitly does not want images, or the core of the task is map display, real attachment generation, ordinary rewriting and translation, or short question answering that requires no graphics.

Core workflow

1. Determine benefit: Trigger by default when the user explicitly requests visualization; also trigger when the user does not say so explicitly, but the task involves trends, proportions, visual evidence, dynamic changes, or complex knowledge structures and graphical representation clearly adds value. When text is already sufficient, answer with text only.
2. Determine goal: From 数据分析、知识讲解、视觉证据、动态探索 choose one primary goal, and optionally add one truly necessary auxiliary goal.
3. Inventory materials: Identify structured data, user original images, textual materials, and verified materials. When user original images are involved, you must determine whether it is 保持原图 or 仅作生成参考。
4. Choose presentation: In ECharts、原图静态叠加、HTML/SVG 交互、生成式知识配图、纯文字 choose the minimum but sufficient modules. When the user explicitly specifies a valid format, prioritize complying with it.
5. Pass the loading gate: First read references/routing.md to complete routing; after determining the presentation, you must fully read the corresponding "mode file group" below before starting to write the option, HTML, process, or image Prompt. Do not generate without reading the corresponding deep specification. Do not unconditionally read all four groups of files.
6. Prepare facts: Real data, years, people, events, and medical or engineering details must come from user materials or legitimate verification results. When verification is impossible, degrade to an empty state, template, data retrieval plan, or clearly labeled example; do not fabricate.
7. Generate and check: Output according to the selected mode, and check accuracy, mobile readability, resource availability, interaction degradation, and correspondence between graphics and text. Only retry or fix failed modules.

Quick routing

用户请求
├─ 核心是准确数值、趋势、占比、排行或分布 → ECharts
├─ 核心证据来自用户原图
│  ├─ 原图必须保持不变
│  │  ├─ 只需指出位置、区域、路径或匹配 → 原图静态叠加
│  │  └─ 需要点击、切换、播放或联动解释 → 原图 + HTML/SVG 交互
│  └─ 允许解释性重绘 → 原图作为参考，生成知识配图
├─ 核心是参数变化、算法步骤、状态迁移或动态几何 → HTML/SVG 交互
├─ 核心是概念、机制、流程、历史、角色关系或作品解读 → 生成式知识配图
└─ 可视化收益不足 → 纯文字

The above branches can be combined. For example, "click parts based on a device photo to view principles" uses 保持原图 + HTML 交互; "redraw a product photo into a cross-sectional knowledge diagram" uses 原图作参考 + 生成式配图; "explain structure from a photo and show parameter changes" can use 原图交互 + ECharts。

Routing hard rules

• User format takes priority: When explicitly requesting an ECharts option, annotation on an original image, interactive demonstration, or generated illustration, prioritize the corresponding valid format.
• Evidence must not be redrawn: When the conclusion depends on the real positions, quantities, boundaries, text, or paths in the user's original image, keep the original image; generated images must not be used to replace evidence.
• Data must not be generated as images: Precise values, ratios, statistical relationships, coordinates, and verifiable trends must not be delegated to generative images.
• Generated images are not proof: Images generated with reference to the original image must be called explanatory redrawings or schematic diagrams, not original images, official images, precise structural diagrams, or factual evidence.
• Interaction must have causal value: Use interaction only when clicking, dragging, playing, switching, or parameter changes can increase understanding; do not package static content as meaningless buttons.
• Combination is allowed but quantity must be controlled: By default, one main module; for complex tasks, add at most the necessary auxiliary modules. Each module must express different information.
• Maps are prohibited: Do not generate maps, administrative divisions, latitude-longitude points, geographic trajectories, tile maps, or ECharts geo/map; instead use text, lists, non-map flows, or data retrieval solutions.
• Attachments are handled separately: Delivery of real attachments such as PDF, PPT, Word, Excel, and image files must use the corresponding file capabilities; do not use the renderer to impersonate files.

Presentation modes and required files

Mandatory progressive loading gate: The file groups in the table are not supplementary materials, but the execution specifications for the corresponding branches. After a branch is matched, all files in that row must be read in full; for combined output, read all matched rows, then read references/composition.md. When only making routing decisions, first read references/routing.md, and do not load irrelevant branches in advance.

Presentation mode	Entry condition	Output contract	File groups that must be read in full
ECharts	Precise data charts, or the user explicitly requests option	Markdown + a single complete echarts option	references/mode-echarts.md → references/echarts-option-spec.md → references/shared-quality.md
Static overlay on the original image	The original image is evidence and only requires points, boxes, lines, paths, matching	Text answer + original image overlay renderer at the end	references/mode-image-overlay.md → references/image-overlay-process-spec.md → references/image-overlay-authoring-spec.md → references/shared-quality.md
HTML/SVG interaction	Requires dynamic exploration; may reference the original image, data, or self-drawn structures	Text explanation + operable renderer	references/mode-interactive.md → references/renderer-trigger-design.md → references/renderer-stability-math.md → references/renderer-interaction-geometry.md → references/renderer-output-mobile.md → references/shared-quality.md
Generative knowledge illustration	Explain concepts or relationships; the original image may be used as reference rather than evidence	Segmented body text + corresponding images	references/mode-generated-illustration.md → references/generated-prompt-rules.md → references/generated-style-guide.md → references/generated-tool-contracts.md → references/shared-quality.md
Combined output	A single mode cannot fully express it, and the modules are complementary	Combine in reading order, do not repeat information	references/composition.md

Loading completion check

Before starting generation, confirm internally:

route_decided = true
required_files_loaded = true
output_contract_fixed = true

If any item is false, continue routing or reading files, and do not start generation. When tool parameters are uncertain, additionally read references/tool-contracts.md; when an implementation reference is needed, then read the corresponding example; examples cannot replace specification files.

Internal planning

Form a brief plan before execution, and do not show internal fields to the user:

{
  "should_visualize": true,
  "goals": ["knowledge_explanation", "dynamic_exploration"],
  "assets": ["user_image", "verified_text"],
  "user_image_policy": "preserve",
  "presentations": ["interactive_html"],
  "required_files": [
    "references/mode-interactive.md",
    "references/renderer-trigger-design.md",
    "references/renderer-stability-math.md",
    "references/renderer-interaction-geometry.md",
    "references/renderer-output-mobile.md",
    "references/shared-quality.md"
  ],
  "required_files_loaded": true,
  "facts_status": "verified",
  "reason": "点击原图部件能直接解释工作原理"
}

See the field values and combination constraints in references/routing.md. Do not automatically use original image annotation just because a user image exists; first determine whether the image is evidence, a base map, or a generation reference.

Unified output requirements

• First give the user the conclusion they most need, then give visualizations and necessary interpretation.
• Explain the data source, factual basis, example attributes, or the nature of the generated image.
• Original-image overlay by default places the visualization after the text answer; knowledge illustrations should immediately follow the corresponding paragraph; only necessary explanations should be kept before and after data charts.
• Combined tasks are organized in the reading order of "conclusion → raw evidence → precise data → explanatory diagrams"; not all are required every time.
• When tools or resources are unavailable, keep the text answer and explain what is missing; do not bypass official tools, construct authentication on your own, or guess private interfaces.

Security and factual boundaries

• Do not fabricate statistics, stock prices, financial reports, medical cases, rankings, years, personal identities, event sequences, or sources.
• Do not echo tokens, cookies, AK/SK, JWTs, private keys, or unrelated personal information in outputs, prompts, HTML, or logs.
• When the user's original image contains sensitive information, annotate only the areas needed to complete the task; do not repeat unrelated phone numbers, ID numbers, addresses, or chat content.
• When search images are used as references for real entities, distinguish search images, the user's original images, and generated images; do not describe search results as your own or official materials.
• High-risk diagrams in medicine, engineering, finance, and similar fields can only assist explanation and do not replace professional judgment; do not let diagrams imply unverified conclusions.

Reference files

• references/routing.md: goals, material strategy, presentation methods, combined routing, and degradation conditions; required reading each time.
• references/shared-quality.md: fact, mobile, resource, interaction, and output quality checks; required reading each time.
• references/mode-echarts.md: rules for precise data charts.
• references/mode-image-overlay.md: rules for static evidence annotation that preserves the original image.
• references/mode-interactive.md: interaction rules for pure self-drawn or original-image-based HTML/SVG.
• references/mode-generated-illustration.md: rules for generative knowledge illustrations based on plain text materials or original-image references.
• references/echarts-option-spec.md: complete option, mobile, free-layout, and template details of the original ECharts Skill.
• references/image-overlay-process-spec.md、references/image-overlay-authoring-spec.md: complete process, coordinates, two-layer stage, visual, and interaction details of original-image annotation.
• references/renderer-trigger-design.md、references/renderer-stability-math.md、references/renderer-interaction-geometry.md、references/renderer-output-mobile.md: complete trigger, design, stability, formula, interaction, geometry, and mobile rules of the original general renderer.
• references/generated-prompt-rules.md、references/generated-style-guide.md、references/generated-tool-contracts.md: complete prompt, style, and tool rules of the original knowledge illustrations.
• references/composition.md: multi-mode combination, order, deduplication, and conflict handling.
• references/tool-contracts.md: available tool capabilities, parameter verification, and handling when tools are unavailable.
• examples/routing-cases.md: positive examples, negative examples, boundary examples, and combined scenarios.
• examples/image-overlay-gold-process.md、examples/image-overlay-gold-reply.md: complete process and renderer implementation of original-image annotation.
• examples/generated-worked-example.md: complete three-image planning and invocation examples for knowledge illustrations.
• references/echarts-source.md: explanation of the original source of the ECharts rules.
• references/migration-coverage.md: migration destinations, retention status, and targeted revision notes for the 18 files across the four original Skills.
• schemas/visualization-plan.schema.json: machine-checkable schema for the internal planning structure.
• scripts/validate_skill.py: check the Skill structure, links, routing keywords, and known conflicts.