scientific-figure-prompt

name	scientific-figure-prompt
description	Use when generating AI drawing prompts for scientific figures, academic papers, or technical presentations. Triggers on requests for 配图、流程图、架构图、示意图、figure prompt、scientific figure、publication-ready illustration.
user-invocable	true

Scientific Figure Prompt

Overview

从技术描述直接生成结构化、出版级质量的 AI 绘图提示词，适配 Midjourney / DALL-E / Gemini。

核心机制：

• 交互式风格选择：7 个预设方案 + 自定义
• 全局风格文档 (global_style.md)：一次生成，全篇论文多图复用，保证视觉一致性
• 三阶段内部推理：Understand → Think → Design — 用户只看到最终 prompt

When to Use

Use when:

• 用户需要为科学论文、技术报告、学术汇报生成配图 prompt
• 用户提到：配图、流程图、架构图、示意图、figure prompt、scientific figure
• 用户需要多张配图共享统一视觉风格

When NOT to use:

• 纯艺术创作、非技术类插图、照片处理

Core Pattern

Step 1: 收集用户输入

用 clarify 依次获取以下信息（可合并为一次询问）：

字段	必填	说明
visualization_goal	Y	这张图要讲什么故事？一句话核心信息
technical_content	Y	要可视化的技术内容：系统/算法/概念，关键组件、关系、数据流
scene_description	Y	大致场景描述：几个模块、大致布局方向、重点突出什么
language	Y	标注语言：中文 / 英文 / 中英双语
global_style	N	全局绘图风格 md（已有则粘贴；无则留空，系统会生成）
special_requirements	N	特殊约束：必须突出 X、避免 Y、类似某论文风格等

如果用户消息中已包含部分信息，只需补充缺失项。

Step 2: 全局绘图风格处理

这是多图风格一致性的核心机制。一篇论文/汇报通常需要多张配图，必须共享同一套视觉语言。

判断逻辑

IF 用户提供了 global_style (完整的 md 格式全局风格文档):
    → 直接使用，跳过风格选择，进入 Step 4

ELIF 用户提供了模糊风格要求 (如 "蓝色系"、"偏简约"):
    → 基于模糊要求 + 默认预设 P1 生成 global_style.md
    → 输出 global_style.md 给用户保存
    → 进入 Step 4

ELSE (用户未提供任何风格信息):
    → 进入 Step 3 风格选择
    → 选择后生成 global_style.md
    → 输出 global_style.md 给用户保存
    → 进入 Step 4

global_style.md 格式

# Global Figure Style — [项目名/论文名]

## Color Palette
- Primary (#XXXXXX): [用途说明，如 "核心模块/主流程"]
- Secondary (#XXXXXX): [用途说明，如 "辅助组件/输出"]
- Accent (#XXXXXX): [用途说明，如 "创新点/重点高亮"]
- Neutral (#XXXXXX): [用途说明，如 "连接线/标注文字"]
- Background (#XXXXXX): [背景色]

## Typography
- Language: [Chinese / English / Bilingual]
- Font: sans-serif (e.g., Inter, Noto Sans)
- Label hierarchy: Module name (14px bold) > Sub-label (12px regular) > Annotation (10px light)

## Shapes & Strokes
- Module: rounded rectangle, 8px radius, 1.5px stroke
- Group boundary: dashed rectangle, 1px stroke, 6px dash
- Arrow: solid, 1.5px, with arrowhead
- Feedback arrow: curved, dashed, 1px
- Icon: minimal line-style, 24px

## Visual Rules
- Shadows: none
- Gradients: none
- 3D effects: none
- Decorative elements: none
- Background: solid color only

## Technical Specs
- Resolution: 4K (3840x2160), 300 DPI
- Default aspect ratio: 16:9 landscape
- Quality: IEEE/ACM publication standard

## Negative Prompt
no watermarks, no blurry, no cluttered, no 3D, no gradients, no shadows, no decorative elements, no stock photo style, no text errors, no overlapping labels, no dark background

使用方式

• 第一次调用：无 global_style → 系统生成并输出 → 用户保存为 figure_style.md
• 后续调用：用户粘贴 figure_style.md 内容到 global_style 字段 → 系统直接使用，不再输出风格文档
• 跨图一致性：同一项目所有图共享同一份 global_style，确保色板、字体、形状、规格完全一致

Step 3: 风格选择（交互式）

仅在 Step 2 中用户未提供 global_style 时执行。

用 clarify 提供预设方案，同时允许用户自定义。选择后自动生成 global_style.md 并输出给用户保存。

Quick Reference

7 种预设方案

编号	预设名	适用场景	关键特征
P1	AI/SE 学术白底简约	论文/汇报配图（默认）	白底、浅色系、无渐变、圆角矩形、1.5px 描边、无 3D、无阴影、sans-serif
P2	Nature/Science 深色	顶刊 Figure	浅灰底(#F5F5F5)、蓝绿主色系、细线连接、subfigure 编号
P3	系统架构图	微服务/系统设计	分层布局、模块色块区分、虚线边界、数据流箭头
P4	数据流水线	ETL/Pipeline/Workflow	左→右流向、阶段色带、图标+标签、并行分支
P5	循环/反馈图	训练循环/迭代优化	环形布局、反馈箭头、收敛指示、中心核心
P6	网络拓扑	分布式系统/通信图	节点-边模型、辐射/网状、协议标注、层次分组
P7	自定义	用户完全自定义	用户描述风格偏好

默认选 P1（AI/SE 学术白底简约）。

预设参数映射

P1: AI/SE 学术白底简约（默认）

background: "#FFFFFF"
color_palette:
  primary: "#4A90D9"
  secondary: "#7EC8A0"
  accent: "#E8855B"
  neutral: "#8C8C8C"
  background: "#FFFFFF"
style:
  medium: clean vector illustration, flat design
  shapes: rounded rectangles (8px radius), 1.5px stroke
  icons: minimal line-style, 24px
  shadows: none
  gradients: none
  effects_3d: none
  typography: sans-serif, clear hierarchy
tech_specs:
  resolution: 4K (3840x2160), 300 DPI
  aspect_ratio: 16:9 landscape
  quality: IEEE/ACM publication standard
negative_prompt: no watermarks, no blurry, no cluttered, no 3D, no gradients, no shadows, no decorative elements, no stock photo style, no text errors, no overlapping labels, no dark background

P2: Nature/Science 深色

background: "#F5F5F5"
color_palette:
  primary: "#2B6CB0"
  secondary: "#38A169"
  accent: "#D69E2E"
  neutral: "#718096"
  background: "#F5F5F5"
style:
  medium: clean vector, subtle line weight variation
  shapes: rounded rectangles (4px radius), 1px stroke
  subfigure_numbering: true
  shadows: none
  gradients: none
tech_specs:
  resolution: 4K, 300 DPI
  aspect_ratio: "3:2"
  quality: Nature/Science publication standard
negative_prompt: no watermarks, no blurry, no 3D, no gradients, no decorative, no text errors

P3: 系统架构图

background: "#FFFFFF"
color_palette:
  primary: "#3182CE"
  secondary: "#805AD5"
  accent: "#DD6B20"
  neutral: "#A0AEC0"
  background: "#FFFFFF"
style:
  medium: clean vector, layered layout
  shapes: rounded rectangles, dashed boundaries for groups
  connections: solid arrows for data flow, dashed for dependencies
  shadows: none
  gradients: none
tech_specs:
  resolution: 4K, 300 DPI
  aspect_ratio: 16:9
  quality: IEEE publication standard
negative_prompt: no watermarks, no blurry, no 3D, no gradients, no decorative, no text errors

P4: 数据流水线

background: "#FFFFFF"
color_palette:
  primary: "#3B82F6"
  secondary: "#10B981"
  accent: "#F59E0B"
  neutral: "#6B7280"
  background: "#FFFFFF"
style:
  medium: clean vector, left-to-right flow
  shapes: rounded rectangles with stage color bands
  connections: solid arrows, parallel branches
  shadows: none
  gradients: none
tech_specs:
  resolution: 4K, 300 DPI
  aspect_ratio: 16:9 landscape
  quality: IEEE publication standard
negative_prompt: no watermarks, no blurry, no 3D, no gradients, no decorative, no text errors

P5: 循环/反馈图

background: "#FFFFFF"
color_palette:
  primary: "#6366F1"
  secondary: "#14B8A6"
  accent: "#EF4444"
  neutral: "#94A3B8"
  background: "#FFFFFF"
style:
  medium: clean vector, circular/radial layout
  shapes: rounded rectangles, curved feedback arrows
  shadows: none
  gradients: none
tech_specs:
  resolution: 4K, 300 DPI
  aspect_ratio: 1:1 square
  quality: IEEE publication standard
negative_prompt: no watermarks, no blurry, no 3D, no gradients, no decorative, no text errors

P6: 网络拓扑

background: "#FFFFFF"
color_palette:
  primary: "#2563EB"
  secondary: "#7C3AED"
  accent: "#DC2626"
  neutral: "#9CA3AF"
  background: "#FFFFFF"
style:
  medium: clean vector, node-edge graph
  shapes: circles for nodes, lines for edges, grouped by hierarchy
  shadows: none
  gradients: none
tech_specs:
  resolution: 4K, 300 DPI
  aspect_ratio: 16:9
  quality: IEEE publication standard
negative_prompt: no watermarks, no blurry, no 3D, no gradients, no decorative, no text errors

P7: 自定义

由用户描述风格偏好，AI 衡量后生成参数集。需确保输出格式与上述预设一致。

Implementation

Step 4: 三阶段推理生成 Prompt

内部执行，不输出中间过程。

Phase 1: UNDERSTAND

• 提取 core story（输入→处理→输出？对比？层次？循环？）
• 列出 key entities 和 relationships
• 识别 innovation/highlight 点
• 确定受众（reviewers / general / technical）和载体（paper / slides / poster）

Phase 2: THINK

• 从 7 种图类型中选择最匹配的，给出理由： System Architecture / Flowchart-Pipeline / Layered / Cycle-Loop / Comparison / Radial-HubSpoke / Swimlane
• 设计视觉层级：FIRST（最大/最醒目）→ SECOND（支撑）→ de-emphasize（必要但不焦点）
• 布局方向：L→R（流程）、T→B（层次）、radial（循环）
• 创新点位置：黄金分割位 / 中心 / 终端

Phase 3: DESIGN

基于 Phase 1-2 结论 + 用户选定预设参数，生成结构化 prompt。

Step 5: 输出格式

情况 A：用户提供了 global_style（后续调用） — 只输出绘图 prompt，不输出 global_style.md。

情况 B：用户未提供 global_style（首次调用） — 先输出 global_style.md（供用户保存），再输出绘图 prompt。

绘图 prompt 格式：

=== SCENE DESCRIPTION ===
[2-3 句话描述：图类型 + 主体 + 核心视觉特征 + 风格]
[核心故事/流向]
[整体构图]

=== VISUAL ELEMENTS ===
[Group 1 Name]:
• [Element]: [shape], [color], [position], [label]

[Group 2 Name]:
• [Element]: [shape], [color], [position], [label]

[Connections]:
• [Source] → [Target]: [arrow style], [label]

=== COMPOSITION & LAYOUT ===
• Direction: [left-to-right / top-to-bottom / radial]
• Focal point: [视觉焦点]
• Grouping: [元素分组方式]
• White space: [留白位置]

=== COLOR PALETTE ===
• Primary (#XXXXXX): [用途]
• Secondary (#XXXXXX): [用途]
• Accent (#XXXXXX): [用途]
• Neutral (#XXXXXX): [用途]
• Background (#XXXXXX)

=== STYLE SPECIFICATIONS ===
• Medium: [from preset]
• Aesthetic: professional, technical, publication-ready
• Shapes: [from preset]
• Icons: [from preset]
• Typography: [language] labels, sans-serif
• Shadows: none
• Gradients: none

=== TECHNICAL SPECS ===
• Resolution: [from preset]
• Aspect ratio: [from preset]
• Quality: [from preset]

=== NEGATIVE PROMPT ===
[from preset]

Step 6: 自检

输出前内部验证：

• 核心故事是否一眼可读？
• 创新点是否视觉突出？
• 用户所有需求是否覆盖？
• 颜色是否为浅色系、无渐变（P1 默认约束）？
• 标注语言是否匹配用户选择？
• Prompt 是否足够具体可复现？

示例 1：首次调用（无 global_style）

输入：

visualization_goal: 展示 RAG 系统如何通过检索增强减少 LLM 幻觉
technical_content: Query→Embedding→Vector DB检索→Context拼接→LLM生成→验证。关键创新：双路检索（稀疏+稠密）+ 自适应重排序
scene_description: 左到右流水线，中间突出双路检索模块，5-6个组件
language: 英文
global_style: (未提供)
preset: P1

输出（先输出 global_style.md，再输出 prompt）：

# Global Figure Style — RAG System Paper

## Color Palette
- Primary (#4A90D9): core modules / main pipeline stages
- Secondary (#7EC8A0): output / generation components
- Accent (#E8855B): innovation highlights / key differentiator
- Neutral (#8C8C8C): connection lines / annotations
- Background (#FFFFFF): white

## Typography
- Language: English
- Font: sans-serif (Inter)
- Label hierarchy: Module name (14px bold) > Sub-label (12px regular) > Annotation (10px light)

## Shapes & Strokes
- Module: rounded rectangle, 8px radius, 1.5px stroke
- Group boundary: dashed rectangle, 1px stroke, 6px dash
- Arrow: solid, 1.5px, with arrowhead
- Feedback arrow: curved, dashed, 1px
- Icon: minimal line-style, 24px

## Visual Rules
- Shadows: none
- Gradients: none
- 3D effects: none
- Decorative elements: none
- Background: solid color only

## Technical Specs
- Resolution: 4K (3840x2160), 300 DPI
- Default aspect ratio: 16:9 landscape
- Quality: IEEE/ACM publication standard

## Negative Prompt
no watermarks, no blurry, no cluttered, no 3D, no gradients, no shadows, no decorative elements, no stock photo style, no text errors, no overlapping labels, no dark background

---

=== SCENE DESCRIPTION ===
A left-to-right pipeline diagram showing a Retrieval-Augmented Generation (RAG) system, with the dual-path retrieval module visually emphasized at the center.
The diagram illustrates how user queries flow through embedding, dual-path retrieval (sparse + dense), adaptive reranking, context assembly, and LLM generation with verification.
Clean flat design on white background, with the innovation module highlighted in warm orange.

=== VISUAL ELEMENTS ===
Input Stage:
• User Query: rounded rectangle, #4A90D9, leftmost, "Query"
• Query Embedding: rounded rectangle, #4A90D9, second, "Embedding"

Retrieval Stage (Highlighted):
• Sparse Retriever: rounded rectangle, #E8855B, center-upper, "Sparse (BM25)"
• Dense Retriever: rounded rectangle, #E8855B, center-lower, "Dense (Vector)"
• Adaptive Reranker: rounded rectangle, #E8855B, center-right, "Adaptive Reranking"

Generation Stage:
• Context Assembly: rounded rectangle, #7EC8A0, right-center, "Context Assembly"
• LLM Generator: rounded rectangle, #7EC8A0, right, "LLM Generation"
• Verifier: rounded rectangle, #7EC8A0, rightmost, "Verification"

Connections:
• Query → Embedding: solid arrow, ""
• Embedding → Sparse Retriever: solid arrow, "sparse path"
• Embedding → Dense Retriever: solid arrow, "dense path"
• Sparse Retriever → Reranker: solid arrow, ""
• Dense Retriever → Reranker: solid arrow, ""
• Reranker → Context Assembly: solid arrow, "ranked docs"
• Context Assembly → LLM Generator: solid arrow, "augmented prompt"
• LLM Generator → Verifier: solid arrow, "output"

=== COMPOSITION & LAYOUT ===
• Direction: left-to-right
• Focal point: dual-path retrieval cluster (center, orange accent)
• Grouping: Input (blue) | Retrieval (orange, highlighted) | Generation (green)
• White space: generous padding between stages, 20px between groups

=== COLOR PALETTE ===
• Primary (#4A90D9): input stage modules
• Secondary (#7EC8A0): generation stage modules
• Accent (#E8855B): retrieval innovation (dual-path + reranker)
• Neutral (#8C8C8C): connection lines, labels
• Background (#FFFFFF)

=== STYLE SPECIFICATIONS ===
• Medium: clean vector illustration, flat design
• Aesthetic: professional, technical, publication-ready
• Shapes: rounded rectangles (8px radius), 1.5px stroke
• Icons: minimal line-style, 24px
• Typography: English labels, sans-serif, clear hierarchy
• Shadows: none
• Gradients: none

=== TECHNICAL SPECS ===
• Resolution: 4K (3840x2160), 300 DPI
• Aspect ratio: 16:9 landscape
• Quality: IEEE/ACM publication standard

=== NEGATIVE PROMPT ===
no watermarks, no blurry, no cluttered, no 3D, no gradients, no shadows, no decorative elements, no stock photo style, no text errors, no overlapping labels, no dark background

示例 2：后续调用（有 global_style）

输入：

visualization_goal: 展示 Verifier 模块如何检测并过滤幻觉内容
technical_content: LLM输出→事实校验(知识库比对)→置信度评分→阈值过滤→高置信输出/低置信回退。关键：多粒度校验（句子级+实体级）
scene_description: 分支结构，上方是正常输出路径，下方是回退路径，中间是校验模块
language: 英文
global_style: (粘贴之前生成的 Global Figure Style — RAG System Paper)

输出（只输出 prompt，不重复输出 global_style.md）：

=== SCENE DESCRIPTION ===
A branching verification diagram showing how the Verifier module detects and filters hallucinated content, with the multi-granularity check visually emphasized.
The diagram illustrates LLM output flowing through fact verification (KB comparison), confidence scoring, and threshold-based routing — high-confidence passes through, low-confidence triggers fallback.
Clean flat design on white background, verification module highlighted in warm orange, consistent with project style.

=== VISUAL ELEMENTS ===
Verification Core (Highlighted):
• Fact Verifier: rounded rectangle, #E8855B, center, "Fact Verification (KB)"
• Sentence-level Check: rounded rectangle, #E8855B, center-upper, "Sentence-level"
• Entity-level Check: rounded rectangle, #E8855B, center-lower, "Entity-level"
• Confidence Scorer: rounded rectangle, #E8855B, center-right, "Confidence Scoring"

High-confidence Path:
• Threshold Gate (pass): diamond, #4A90D9, right, "Score >= theta"
• Final Output: rounded rectangle, #7EC8A0, far-right, "High-confidence Output"

Low-confidence Path:
• Threshold Gate (fail): diamond, #8C8C8C, below-right, "Score < theta"
• Fallback: rounded rectangle, #8C8C8C, bottom-right, "Fallback / Re-query"

Connections:
• LLM Output → Fact Verifier: solid arrow, ""
• Fact Verifier → Sentence-level Check: solid arrow, ""
• Fact Verifier → Entity-level Check: solid arrow, ""
• Sentence-level Check → Confidence Scorer: solid arrow, ""
• Entity-level Check → Confidence Scorer: solid arrow, ""
• Confidence Scorer → Threshold Gate (pass): solid arrow, ">= theta"
• Confidence Scorer → Threshold Gate (fail): solid arrow, "< theta"
• Threshold Gate (pass) → Final Output: solid arrow, ""
• Threshold Gate (fail) → Fallback: solid arrow, ""

=== COMPOSITION & LAYOUT ===
• Direction: left-to-right with vertical branch at threshold
• Focal point: multi-granularity verification cluster (center, orange accent)
• Grouping: Verification (orange) | Pass path (blue→green) | Fallback path (gray)
• White space: 20px between groups, branch fork clearly separated

=== COLOR PALETTE ===
• Primary (#4A90D9): threshold gate (pass)
• Secondary (#7EC8A0): final output
• Accent (#E8855B): verification core (multi-granularity)
• Neutral (#8C8C8C): fallback path, threshold gate (fail)
• Background (#FFFFFF)

=== STYLE SPECIFICATIONS ===
• Medium: clean vector illustration, flat design
• Aesthetic: professional, technical, publication-ready
• Shapes: rounded rectangles (8px radius), 1.5px stroke; diamonds for decision
• Icons: minimal line-style, 24px
• Typography: English labels, sans-serif, clear hierarchy
• Shadows: none
• Gradients: none

=== TECHNICAL SPECS ===
• Resolution: 4K (3840x2160), 300 DPI
• Aspect ratio: 16:9 landscape
• Quality: IEEE/ACM publication standard

=== NEGATIVE PROMPT ===
no watermarks, no blurry, no cluttered, no 3D, no gradients, no shadows, no decorative elements, no stock photo style, no text errors, no overlapping labels, no dark background

Common Mistakes

错误	修正
使用深色/渐变/3D 效果	P1 默认禁止；白底、浅色系、flat design
输出 Phase 1/2 推理过程	用户只要最终 prompt，不输出中间推理
让用户手动填模板文件	直接交互式收集信息
生成模糊色值（如 "light blue"）	必须给 hex 值（如 #4A90D9）
遗漏 negative prompt	每次输出必须包含
用户提供了 global_style 时重新生成风格文档	直接使用，不重新生成
首次调用跳过 global_style.md 输出	未提供时必须生成并输出
global_style.md 留占位符	所有值必须填充具体值
多轮追问同一信息	尽量一次 clarify 搞定