Auto Paper Figure Skill / 科研论文插图自动绘制 Skill

A reusable GitHub-ready skill for automatically generating paper overview / method / system figures from a real codebase and local paper references.

这是一个可直接放到 GitHub 的 skill 项目，目标是：先理解真实项目实现，再自动生成 ACL / EMNLP 风格的科研论文图片，尤其适用于 overview figure、method figure、system figure，以及包含 LLM harness、training pipeline、核心模块与关键公式 的论文图。

---

What this repo contains / 仓库内容

• An optimized general prompt distilled from a real successful workflow, with project-specific details removed.
• Chinese and English skill versions that can be used as reusable instruction files.
• A README template and usage guide for turning the skill into a portable GitHub project.
• A reference image showing the desired visual direction.

---

Target capability / 目标能力

This skill is designed for the following workflow:

1. Read the implementation first rather than hallucinating a pipeline.
2. Identify the true method narrative, especially the harness and training flow.
3. Inspect local reference papers and imitate their design language rather than copying their content.
4. Draw figures in an ACL/EMNLP-style scientific illustration aesthetic:
   • fewer words
   • clear hierarchy
   • less like a generic flowchart
   • light comic / hand-drawn module styling
   • preserve key technical details when necessary
5. Export editable and publication-friendly outputs, ideally SVG + PDF, plus PNG previews.

---

Reference figure / 参考风格图

Below is the reference image included in this repo:

---

Repository structure / 目录结构

paper-overview-figures-skill/
├── README.md
├── .gitignore
├── LICENSE
├── reference_overview_example.png
├── prompts/
│   ├── optimized_general_prompt_zh.md
│   └── optimized_general_prompt_en.md
├── skills/
│   ├── paper_overview_figures_skill_zh.md
│   └── paper_overview_figures_skill_en.md
├── examples/
│   ├── example_usage_zh.md
│   └── example_usage_en.md

---

Recommended usage / 推荐使用方式

Option A: use the optimized prompt directly / 直接使用优化后的 prompt

Use the prompt under prompts/ when you want an agent to inspect a project and generate a paper overview figure.

适合你已经有自己的 agent / coding assistant / skill runtime，只需要一份高质量通用 prompt 的情况。

Option B: use the skill files / 使用 skill 文件

Use the skills/ files when your environment supports skill-style reusable instruction documents.

适合你希望把这个能力封装成长期复用的“画科研图 skill”的情况。

---

Design principles / 设计原则

1) Accuracy first / 科学准确性优先

The figure should reflect the actual implementation, not a guessed workflow.

2) Narrative over call graph / 方法叙事优先于代码调用链

The goal is not to dump boxes and arrows, but to tell the paper story.

3) Minimal text, not minimal meaning / 少文字，不少信息

Remove redundant explanation, but keep the details that define the method.

4) Scientific illustration, not business flowchart / 做科研图，不做业务流程图

Prefer structured scientific illustration with visual metaphors, grouped modules, and focused emphasis.

5) Controlled hand-drawn aesthetics / 克制的手绘感

Use light comic / hand-drawn styling to make the figure more natural, without reducing readability.

---

Files you will likely edit / 你最可能修改的文件

• prompts/optimized_general_prompt_zh.md General Chinese prompt distilled from the original real usage.
• prompts/optimized_general_prompt_en.md English counterpart.
• skills/paper_overview_figures_skill_zh.md Chinese reusable skill.
• skills/paper_overview_figures_skill_en.md English reusable skill.

---

Suggested downstream outputs / 建议下游输出

When this skill is used on a project, a good default deliverable set is:

• figure/overview_en.pdf
• figure/overview_zh.pdf
• figure/overview_en.svg
• figure/overview_zh.svg
• figure/overview_en.png
• figure/overview_zh.png
• figure/src/ for editable sources or generation scripts
• figure/README.md documenting evidence, design choices, and export steps

---

Example request summary / 示例请求概括

A typical request looks like this:

• Read my project implementation.
• Focus on the LLM harness and training pipeline.
• Draw an ACL/EMNLP-style overview figure.
• Reduce flowchart feel, reduce text density.
• Use light comic-style modules.
• Restore necessary technical details and formulas in key modules.
• Add a natural hand-drawn touch.
• Export PDF/SVG/PNG and keep editable sources.

---

License / 许可

This repository is packaged with an MIT license for convenient reuse.

---

Notes / 说明

This repo contains prompts and skill files, not a fixed rendering engine. It is intended to be plugged into an agentic workflow that can read project files, inspect local paper PDFs, and generate figures.

这个仓库主要提供的是 prompt / skill 资产，并不绑定某一个固定绘图引擎；它更适合接入一个能够读取项目代码、查看本地论文并生成图片的 agent 工作流。