SuperAPI AI Directory
Partner中文AI newsFavorites
Skillspaper-comic

paper-comic

Writing & research

View repository

Updated 10-16 10:04

About this skill

Paper Method Illustrated — Paper Method Illustrated

Thoroughly explain the paper's core method with visual diagrams.

The essential difference from other skills

Traditional paper-comic baoyu-article-illustrator Us: method illustration
What is discussed A little bit of everything Add decorative images to the article Only discuss method details
Depth One page covers a lot → shallow One image and one paragraph → medium One image fully explains one mechanism → deep
Number of pages Fixed at 10 pages 5-20 images depending on density First recommend a combination of cover/overview/detail images; after user confirmation, generate 1-10 images
Focus Discuss background + story Beautify article layout Visualize "how it is done"

Quick start

/paper-comic /path/to/paper.pdf
/paper-comic https://arxiv.org/abs/2512.xxxxx
/paper-comic /path/to/paper.pdf --style sketchnote
/paper-comic /path/to/paper.pdf --style paper-figure --language English --pages 1

Image generation: automatic detection

Like paper-analyzer, do not hard-code any API. Automatically detect at runtime:

Environment Automatically use
Codex Built-in imagegen skill
Claude Code Installed image-generation skill (such as baoyu-image-gen)
Cursor/Other Automatically detect → if not found, prompt installation

Do not hard-code any API key, token, or endpoint in SKILL.md.

Core philosophy

We draw only three things

  1. Method workflow — input → processing → output, how this method actually proceeds
  2. Core mechanism — the most innovative part; open it up and look at its internal structure
  3. Key results — include only the most important experimental results, not filler data

What we do not draw

  • ❌ Related work/background introduction (that is the job of the text part of paper-analyzer)
  • ❌ Abstract "sources of inspiration" (an image with no information = a waste)
  • ❌ Things that text alone can make clear (if one sentence can explain it, there is no need to draw it)
  • ❌ The bar chart of the Nth ablation experiment

The standard for every image

If someone who has never read the paper looks only at this image + the annotation text on it, can they understand this mechanism?

Yes → pass. No → split it into two images, or add more annotations.

Must confirm before generation

By default,do not generate images directly. First read the paper, provide a recommended plan, and then confirm with the user.

Only when the user has already given a sufficiently complete generation intent may confirmation be skipped, for example:

  • "Generate one Chinese sketchnote method overview image"
  • "Generate 4 images: cover, overview, two mechanism details, English paper-figure"
  • "Generate all according to your recommendations, Chinese, sketchnote"

If the user only gives a paper link, only says the style (such as "sketchnote"), or only says "generate an illustration", confirmation is still required, becauseStyle is not the same as scope/number of images authorization。

When confirming, you must cover:

  1. Image language: Chinese / English / bilingual
  2. Generation scope: only a cover image, only a method overview image, or an overview image + several mechanism detail images + result images
  3. Recommended number of images: give a suggestion based on the complexity of the paper, and explain why, e.g. "I suggest 6 images, because this paper has an overall architecture, two core attention mechanisms, encoder/decoder structure, and key experimental results"
  4. Visual style：sketchnote or paper-figure
  5. Purpose: README/article cover/Xiaohongshu/presentation slides/paper reading notes (the purpose determines the aspect ratio and text density)

Example confirmation script:

After reading it, I suggest generating 6 images: 1 cover, 1 method overview, 3 mechanism details, and 1 key result. You can also generate just 1 overview image, or expand to 8 images to explain each mechanism in more detail. Which scope do you want to generate? Should the language be Chinese/English/bilingual? Should the style be sketchnote or paper-figure?

If the user does not answer, do not continue generating.

Two visual styles

Art style Visual effect Suitable scenarios Features
sketchnote(default) Warm research notes style Explaining what the paper is doing, video promotion, knowledge sharing Neat but with a human touch, allowing small symbols, small metaphors, and small visual anchors so people can understand at a glance
paper-figure Paper framework figure style README first screen, paper explanation articles, method overview, technical showcase Like the overview framework figures in top conference papers, but more complete, more beautiful, and more suitable for dissemination

Default recommendation sketchnote. When the user wants "as professional as a paper Figure", "method framework figure", "technical architecture figure", or "very impressive on the first screen of a README", recommend paper-figure。

Detailed sketchnote style specification

  • Bright and warm light beige-white background (close to #FFF8EA / #FAF4E6), like clean handwritten newspaper or classroom handouts
  • Do not use kraft paper, old parchment, vignettes, stains, yellowed edges, or obvious aged textures
  • The main body is black hand-drawn lines + text, with variations in ink thickness
  • Key concepts are emphasized with clean colored pencil/marker texture (dark blue/coral red/olive green/soft yellow), colors are lively but not oversaturated
  • Arrows and connecting lines have a hand-drawn imperfect feel
  • The text is handwritten (English can use a handwritten style; Chinese remains clear and readable)
  • Overall it looks like a bright, warm, information-rich research handwritten poster, not a retro notebook, not a childish comic
  • A small number of fun symbols that aid understanding can be added: magnifying glass, asterisk, sticky note, circled annotation, hand-drawn small light bulb, simplified small icons
  • Fun elements must serve understanding and must not steal attention from the main method figure
  • The main illustration should occupy 75%-85% of the image, avoiding large blank areas; if the page has obvious whitespace, prioritize adding local zoom-in boxes, small examples, dimension annotations, or comparative explanations
  • Each page has a "handwritten" page number in the lower right corner

Detailed paper-figure style specification

  • White background or extremely light gray background, like high-quality method overview figures in NeurIPS / Nature / Science papers
  • Use clean vector-style modules: rounded rectangles, small matrix cells, flow arrows, grouping boxes, numbered steps
  • Restrained but modern color palette: mainly black/dark gray, 1-2 accent colors (choose one or two from blue, cyan, orange, purple)
  • Structure is clearer than the original paper figure: keep the core mechanism, reorganize the layout, avoid copying the original figure
  • Can include small result illustrations, matrix heatmaps, token sequences, module stacking, comparison paths
  • Annotations should be like short labels in paper figure captions: precise, short, professional
  • Suitable for landscape 16:9, 4:3, or portrait 2:3; README first screen prefers landscape or wide images

Workflow

Step 1: Analyze the paper → extract "illustratable content"

After reading the paper, list all content points of the paper, thenkeep only the parts that need illustration：

Must illustrate (1-2 pages each):

  • Overall pipeline/architecture of the method (input → modules → output)
  • Each core innovation mechanism (break it open to show the internals)
  • The most convincing experimental result

Optional illustration (add only if the method is complex):

  • Variants/extensions of the method
  • Key data processing pipeline
  • Visual comparison with baselines

Do not illustrate:

  • Related work (just mention it in text)
  • Multiple similar ablation experiments
  • Background knowledge introduction

Step 2: Give a recommendation and confirm requirements

First output a brief recommendation, do not generate immediately:

我建议生成 6 张：
1. 封面图：论文一句话贡献 + 视觉锚点
2. 方法总览图：解释整体输入、核心模块、输出
3. 核心机制A：拆开最重要的创新点
4. 核心机制B：解释训练/推理/数据流中的关键环节
5. 核心机制C：补足容易误解的内部细节
6. 关键结果图：用一张图说明为什么有效

也可以：
- 只生成 1 张总览图
- 生成 3 张：总览 + 2 张核心机制
- 扩展到 8-10 张，把每个机制讲得更细

请确认：
- 语言：中文 / English / 双语
- 风格：sketchnote / paper-figure
- 范围：只要封面/总览，还是生成全部推荐图？

If the user does not respond, do not continue generating.

Step 3: Determine the number of pages

Based on the paper's complexity, the AI only makes arecommendation, with the final confirmation by the user:

Paper complexity Recommended pages Content allocation
Cover/dissemination image 1 page One cover or high-level overview that clearly explains what the paper does
Quick understanding 2-3 pages Overview + core mechanisms + results
Medium (2 core methods) 4-6 pages Cover/overview + 2-3 mechanisms + key results
Complex (3+ core methods) 6-10 pages Cover/overview + 1 page per mechanism + comparison/results

Rules: Minimum 1 page, maximum 10 pages. Better fewer than more—1 overview image that explains things clearly is better than 10 that explain things confusingly.

Step 4: Write a detailed content description for each page

Do not "generate a prompt"; instead, first use natural language to clearly describewhat exactly this page is meant to express：

第3页：多头注意力机制的内部构造

这一页要讲清楚：Q、K、V是怎么算出来的，它们之间怎么交互。

画面布局（从左到右）：
- 左侧：输入x，一个向量表示
- 中间上方：三条线分别到三个方框（Linear_Q, Linear_K, Linear_V）
- 三个方框各产出Q、K、V三个矩阵
- 中间核心区域：Q和K做点积→除以√dk→softmax→得到注意力权重
- 权重和V相乘→输出
- 右侧：多个这样的"头"并行排列，最后拼接

关键标注：
- 每个方框旁标运算和维度（如"Linear_Q: x→Q(d×dk)"）
- Q×K^T的计算用可视化的矩阵乘法图（小格图）
- softmax后的权重用颜色深浅表示（越深=越关注）

Requirements: The description must be specific enough to say "where this arrow goes from and to, and what text goes in this box."

At the same time, check the information density of each page:

  • If it is only a large title + a few modules, this page is too empty and must be supplemented with small mechanism examples, local zoom-ins, input/output dimensions, or key comparisons
  • If the content exceeds the readable range of one page, split it into two pages; do not cram all the text into the same image
  • The cover image can be more conceptual; mechanism detail images must prioritize clearly explaining "how it is done"

Step 5: Generate images

Automatically select the image generation backend based on the current runtime environment. Create a prompt file for each page → generate with a structured prompt.

Structured prompt format(refer to but do not copy baoyu):

【类型】流程分解图
【风格】sketchnote
【语言】中文
【主题】多头注意力机制内部构造
【视觉结构】
- 水平布局，从左到右5个区域
- 每个区域用虚线框隔开
- 关键路径用粗箭头连接

【要标注的文字】
1. Input: x ∈ R^(n×d)
2. Q = xW_Q ... (完整标注)
...

【颜色限制】
- 背景：明亮浅米白，不要泛黄旧纸
- 主色：黑色手绘线条
- 强调色：深蓝/珊瑚红/橄榄绿/柔和黄色，少量使用
- 其他：保持清爽手抄报感，避免复古暗色

【禁止】
- 不要代码块
- 不要照片写实
- 不要3D渲染
- 不要生成用户没有确认的额外页面
- 不要旧羊皮纸、暗角、污渍、重纸纹、黄褐色复古调
- 不要大面积空白；主体图解占画面75%-85%

Step 6: Output

Generate [topic]-illustrated.md：

# [论文标题] — 图解

## 论文信息
- 论文：[链接]
- 风格：sketchnote
- 页数：6

## 封面
![封面](00-cover.png)
**一句话**：[论文做了什么，为什么重要]

## 第1页：方法总览
![方法流程](01-method-overview.png)
**讲解**：整个方法从输入到输出的完整流程。关键看第X步，这是本文的创新。

## 第2页：核心机制A — [名称]
![核心机制A](02-mechanism-a.png)
**讲解**：这个机制解决了XX问题。具体做法是...关键设计在于...

[重复...]

## 总结：3个核心要点
1. [要点1]
2. [要点2]  
3. [要点3]

Quality standards

Good diagrams

  • ✅ At a single glance, you can tell what the image is about
  • ✅ Label text is precise, concise, and not wordy
  • ✅ Flow arrows are clear, with an explicit "from here to there"
  • ✅ Key parts are visually emphasized with color/size
  • ✅ Understandable without reading the original paper

Bad diagrams

  • ❌ Information is piled up, trying to draw everything in
  • ❌ Too much text, and the image becomes decoration
  • ❌ The flow is unclear, and you don't know what to look at first and what next
  • ❌ Drawn but not explained—an architecture diagram was included but key points were not labeled
  • ❌ It is exactly the same as Figure 1 in the paper—then what is the point of what you drew

Reference files

  • references/base-prompt.md — Basic specifications for diagram generation (structure, text, color requirements)
  • references/styles/sketchnote.md — Warm research note style
  • references/styles/paper-figure.md — Paper framework diagram style