A skill for generating publication-ready scientific figures from AI/CS research papers.
Classify paper type first, then select the optimal diagram structure automatically.
点击展开 / 收起
CCF Figure 是一个专门为 AI / 计算机科学领域的研究人员设计，帮助你将论文内容转化为顶会顶刊级科研配图的工具。
核心理念来自「深度之眼」2026年6月11日发布的文章《科研绘图提示词第二弹》：
先判断论文类型，再选择最合适的图示结构，而不是千篇一律地套用“左输入-中模型-右输出”三段式模板。
| 平台 | 使用方式 |
|---|---|
| Claude Code | 安装到 skills 目录后，在项目中自动调用 |
| Codex | 安装到 skills 目录后，生成完整提示词并辅助调用图像模型 |
- 自动分类：识别 7 种论文类型（方法 / 机制 / 评测 / 规律 / 机器人 / 交叉 / 综述）
- 结构自选：从 11 种图示结构中自动匹配最优布局
- 完整提示词：内置中英文双语提示词模板，可直接复制使用
- 五大翻车预防：内置 Self-Check 清单，避免 AI 绘图的常见错误
- 迭代协议：标准化修改流程，最多 3 轮迭代
| 论文类型 | 推荐图示结构 |
|---|---|
| A · 方法类 | 方法总览图 / 模型架构图 |
| B · 机制 / 分析类 | 机制示意图（局部放大）/ 对比式方法图 |
| C · 评测 / 基准类 | 评测流程图 / 数据构建流程图 |
| D · 规律发现类 | 实验规律总结图 / 实验矩阵图 |
| E · 机器人 / 具身智能 | 闭环反馈图（感知 → 规划 → 执行 → 反馈） |
| F · 交叉方向 | 跨学科应用框架图 |
| G · 综述 / 位置 | 分类树 / 时间线 / 对比矩阵 |
| 属性 | 规则 |
|---|---|
| 背景 | 纯白 #FFFFFF 或极浅灰 #F5F5F5 |
| 风格 | 二维扁平化矢量 — 禁止 3D 透视 / 霓虹渐变 / 装饰纹理 |
| 配色 | 功能性配色；最多 3 个色彩群组；强调色限橙 / 青绿 / 深蓝 |
| 文字 | 水平排布；最多 2 种字号；中文标签 + 英文专有名词 |
| 目标刊会 | NeurIPS, ICML, ICLR, CVPR, ACL, Nature MI, IEEE TPAMI 等 |
# 方式一：安装到 Claude Code 个人级 skills 目录（推荐，所有项目通用）
git clone https://github.com/Deepshare-Official/CCF-Figure \
~/.claude/skills/ccf-figure
# 方式二：安装到当前项目的 Claude Code skills 目录（仅当前项目可用）
git clone https://github.com/Deepshare-Official/CCF-Figure \
.claude/skills/ccf-figure# 方式一：安装到 Codex 个人级 skills 目录（推荐，所有项目通用）
git clone https://github.com/Deepshare-Official/CCF-Figure \
~/.agents/skills/ccf-figure
# 方式二：安装到当前项目的 Codex skills 目录（仅当前项目可用）
git clone https://github.com/Deepshare-Official/CCF-Figure \
.agents/skills/ccf-figure安装完成后，在 Claude Code 或 Codex 中直接说：
帮我为这篇论文生成科研配图：
[发送你的论文原文，或粘贴论文标题 + 摘要 + 方法部分]
CCF Figure 会自动：
- 判断论文类型
- 选择最合适的图示结构
- 组装完整提示词
- 生成图像或输出可直接用于图像模型的完整提示词
CCF-Figure/
├── SKILL.md ← 主工作流（Skill 入口）
├── references/
│ ├── prompt-template.md ← 完整提示词模板（中英双语，可直接复制）
│ └── figure-types.md ← 11 种图示类型规范 + 五大翻车问题
└── README.md ← 本文件
为了让 Claude Code 和 Codex 更稳定识别，建议 SKILL.md 开头使用如下 metadata：
---
name: ccf-figure
description: Generate publication-ready scientific figures for AI and computer science papers. Use when the user provides a paper title, abstract, method section, or full paper and wants a top-conference-style research diagram.
---
# CCF FigureClick to expand / collapse
CCF Figure is a Skill for researchers in AI and computer science.
It transforms paper content into publication-quality scientific figures that meet the visual standards of top conferences and journals.
The core methodology comes from DeepShare’s June 2026 article:
Classify the paper type first, then select the optimal diagram structure — rather than mechanically applying a generic “left-input → center-model → right-output” template.
| Platform | How it works |
|---|---|
| Claude Code | Install it into the skills directory, then invoke it naturally in your project |
| Codex | Install it into the skills directory, then use it to assemble complete scientific-figure prompts |
- Auto-classification: Identifies 7 paper types — method, mechanism, benchmark, scaling, robotics, interdisciplinary, and survey
- Structure auto-selection: Automatically matches the best layout from 11 diagram structures
- Complete prompt library: Bilingual Chinese / English prompt templates, ready to copy and use
- 5-failure-mode prevention: Built-in self-check checklist to avoid common AI figure mistakes
- Iteration protocol: Standardized revision workflow, up to 3 rounds
| Paper Type | Recommended Diagram |
|---|---|
| A · Method | Method Overview / Model Architecture |
| B · Mechanism / Analysis | Mechanism Zoom-in / Contrastive Diagram |
| C · Benchmark / Evaluation | Evaluation Pipeline / Data Construction Pipeline |
| D · Scaling Law / Empirical | Trend Panel / Experimental Matrix |
| E · Robot / Embodied AI | Closed-Loop Feedback Diagram |
| F · Interdisciplinary / AI for X | Cross-domain Application Framework |
| G · Survey / Position | Taxonomy Tree / Timeline / Comparison Matrix |
| Property | Rule |
|---|---|
| Background | Pure white #FFFFFF or very light gray #F5F5F5 |
| Style | 2D flat vector — NO 3D, NO neon gradients, NO decorative textures |
| Color | Functional only; max 3 color groups; accent colors limited to orange, cyan-green, and deep blue |
| Typography | Horizontal only; max 2 font sizes; Chinese labels + English proper nouns |
| Target venues | NeurIPS, ICML, ICLR, CVPR, ACL, Nature MI, IEEE TPAMI, etc. |
# Option 1: Install as a user-level Claude Code skill, available across projects
git clone https://github.com/Deepshare-Official/CCF-Figure \
~/.claude/skills/ccf-figure
# Option 2: Install as a project-level Claude Code skill, available only in the current project
git clone https://github.com/Deepshare-Official/CCF-Figure \
.claude/skills/ccf-figure# Option 1: Install as a user-level Codex skill, available across projects
git clone https://github.com/Deepshare-Official/CCF-Figure \
~/.agents/skills/ccf-figure
# Option 2: Install as a project-level Codex skill, available only in the current project
git clone https://github.com/Deepshare-Official/CCF-Figure \
.agents/skills/ccf-figureAfter installation, simply say in Claude Code or Codex:
Generate a scientific figure for this paper:
[Paste your paper, or paste the paper title + abstract + method section]
CCF Figure will automatically:
- Classify the paper type
- Select the most appropriate diagram structure
- Assemble the complete image-generation prompt
- Generate the figure or output a complete prompt ready for an image model
CCF-Figure/
├── SKILL.md ← Main workflow / Skill entry point
├── references/
│ ├── prompt-template.md ← Full prompt template, bilingual and copy-paste ready
│ └── figure-types.md ← 11 diagram type specs + 5 failure modes
└── README.md ← This file
For better compatibility with Claude Code and Codex, it is recommended to start SKILL.md with the following metadata:
---
name: ccf-figure
description: Generate publication-ready scientific figures for AI and computer science papers. Use when the user provides a paper title, abstract, method section, or full paper and wants a top-conference-style research diagram.
---
# CCF FigureMIT License — feel free to use, modify, and distribute with attribution.
Made with ❤️ for the AI research community · 深度之眼 Deepshare