Paper Figure Gen

学术 Prompt 工程

通用学术约束（EVERY prompt 必须追加）

CRITICAL STYLE REQUIREMENTS:
- Flat vector illustration style, clean lines, minimalist academic aesthetic
- Background: pure white (#FFFFFF), absolutely no texture, gradient, or pattern
- Color palette: muted pastel tones only (soft blue, warm orange, sage green, light gray)
- NO photorealistic rendering, NO 3D shadows, NO glossy/metallic effects
- NO decorative elements, NO ornamental borders, NO watermarks
- Clean geometric shapes: rectangles, rounded rectangles, circles, arrows
- Professional and restrained — this is for a NeurIPS/ICML/CVPR paper, not a poster

类型专用 Prompt 前缀

architecture:

Create a technical architecture diagram for an academic paper.
Layout: clear left-to-right or top-to-bottom data flow.
Components: labeled rectangular modules with short English text inside.
Connections: clean directional arrows showing data/information flow.
Group related components with subtle light-colored background regions.
Hierarchy: main pathway should be visually dominant; auxiliary paths thinner.

pipeline:

Create a horizontal step-by-step pipeline diagram for an academic paper.
Each processing stage is a rounded rectangle with a concise English label.
Stages connected by clean arrows. Use subtle color coding to distinguish phases.
Include small icons or symbols inside stages only if they aid comprehension.
Keep it linear and easy to follow from left to right.

concept:

Create a minimal conceptual illustration for an academic paper.
Use simple abstract geometric shapes to convey the idea.
Maximum clarity, minimum detail. No literal or realistic depictions.
Focus on the core insight or contrast the paper is trying to communicate.
If showing a problem/solution contrast, use clear spatial separation.

comparison:

Create a side-by-side comparison diagram for an academic paper.
Clear visual separation between panels (left vs right, or top vs bottom).
Use consistent visual encoding across panels so differences are obvious.
Label each panel clearly. Highlight key differences with color or annotation.

文字处理策略

Gemini 生成的图内文字经常出现乱码、拼错、缺字问题。采用以下两阶段策略：

阶段 1 — 先尝试带文字生成： 在 prompt 中明确列出所有文字标签：

Text labels that MUST appear exactly as written (spell each correctly):
- "Encoder" (inside the blue rectangle on the left)
- "Attention" (inside the orange rectangle in the center)  
- "Decoder" (inside the green rectangle on the right)

阶段 2 — 如果文字渲染失败，切换为无文字模式： 将 prompt 替换为：

IMPORTANT: Do NOT render any text, letters, words, or characters inside the image.
Leave clean blank spaces where labels would go. Text will be added in post-processing.

然后告知用户需要手动用 draw.io / Figma / PowerPoint 叠加文字标注。

工作流

本 skill 使用 figure-checker subagent 进行自动视觉质量检查。确保该 agent 已配置在 ~/.pi/agent/agents/figure-checker.md。

Step 1: 理解需求

1. 阅读用户提供的论文/方法描述
2. 判断 figure 类型（architecture / pipeline / concept / comparison）
3. 提取需要出现在图中的关键模块名称和连接关系
4. 确认输出尺寸偏好（单栏 0.48\textwidth 还是双栏 0.95\textwidth）

Step 2: 构造 Prompt

按以下顺序拼接 prompt：

[类型专用前缀]

[用户描述的具体内容——模块、流程、关系等]

[文字标签列表]

[通用学术约束]

Step 3: 1K 草稿生成

uv run $GEMINI_SCRIPT \
  --prompt "<constructed prompt>" \
  --filename "figures/yyyy-mm-dd-hh-mm-ss-draft.png" \
  --resolution 1K

Step 4: 强制 Inspect-Revise 循环（最多 3 轮）

生成后，必须 调用 figure-checker subagent 做自动视觉审查，而不是手工逐项检查。

调用格式：

spawn agent=figure-checker task="Check the image at: figures/<draft>.png
Scene: academic
Intent: <constructed prompt summary - what this figure should show>

Additional context:
- Figure type: <architecture|pipeline|concept|comparison>
- Required labels: <list of text labels that should appear>"

根据 figure-checker 返回报告处理：

• ✅ PASS → 直接进入 Step 5
• ⚠️ MINOR ISSUES → 可选地做小修复，然后进入 Step 5
• ❌ REGENERATE → 根据报告中的 Regeneration Guidance 构造修复 prompt，使用 --input-image 编辑模式修复后再次检查

修复命令：

uv run $GEMINI_SCRIPT \
  --prompt "<fix prompt constructed from Regeneration Guidance>" \
  --filename "figures/yyyy-mm-dd-hh-mm-ss-rev1.png" \
  --input-image "figures/<previous-draft>.png" \
  --resolution 1K

常用修复指令（可作为参考）：

• "Make the background pure white. Remove all shadows and gradient effects."
• "Simplify the shapes to flat rectangles. Remove 3D rendering."
• "Fix the text label 'Encoer' — it should be 'Encoder'."
• "Make colors more muted and pastel. Current blue is too saturated."
• "Add clearer arrows between the modules showing data flow direction."

最多迭代 3 轮（生成/检查总轮数）。 如果 3 轮后仍未通过：

1. 保存当前最佳版本
2. 告知用户当前结果及具体不足之处
3. 建议用 draw.io / Figma / TikZ 进行手动微调

Step 5: 4K 终稿

草稿确认后，生成高分辨率终稿：

uv run $GEMINI_SCRIPT \
  --prompt "<same final prompt>" \
  --filename "figures/yyyy-mm-dd-hh-mm-ss-final.png" \
  --resolution 4K

Step 6: LaTeX Snippet 输出

为用户生成可直接粘贴的 LaTeX 代码：

单栏图（concept / comparison）：

\begin{figure}[t]
    \centering
    \includegraphics[width=0.48\textwidth]{figures/<filename>.png}
    \caption{<根据 prompt 内容生成 1-2 句描述性 caption>}
    \label{fig:<auto-label>}
\end{figure}

双栏图（architecture / pipeline）：

\begin{figure*}[t]
    \centering
    \includegraphics[width=0.95\textwidth]{figures/<filename>.png}
    \caption{<caption>}
    \label{fig:<auto-label>}
\end{figure*}

Caption 写作要求：

• 必须 self-contained（不读正文也能理解图的含义）
• 描述图中展示了什么 + 关键发现/结构
• 所有缩写首次出现时展开

输出产物

figures/
├── <timestamp>-draft.png        # 1K 草稿（可删除）
├── <timestamp>-rev1.png         # 修订版（可删除）
├── <timestamp>-final.png        # 4K 终稿 ← 论文使用这个
└── latex-snippets.tex           # LaTeX 引用代码

已知限制

1. 文字渲染不稳定：Gemini 对图内文字的拼写和排版控制有限，复杂标注建议后期手动叠加
2. 几何精度有限：对齐、等距、严格对称等要求难以精确满足
3. 跨图风格一致性：多张图之间的风格统一需要使用 --input-image 以前一张图作为参考
4. 不适用于数据图表：折线图、柱状图、热力图等请用 matplotlib/seaborn

与其他 skill 的关系

设计参考

• Prompt 模板参考 PaperForge prompt-library「论文架构图」部分
• 质量标准参考 DeepScientist ds-figure-polish 的 render-inspect-revise 工作流
• Figure scope 参考 ARIS paper-figure 的 can/cannot 矩阵