linyuebanzi-inline-diagram

Lin Yuebanzi Long-Form Technical Article Illustration Generator

Automatically add 4-5 concept diagrams, flowcharts, and comparison charts to a 3,000-8,000 word long-form technical article, so the article is no longer monotonous.

When to trigger this skill

Explicit triggers:

"Help me add a few images to this article"
"This one is too monotonous, add a few hand-drawn images"
"Add some illustrations, not too many, just at the key positions"
"Take a look at where images could be added"
"Draw a few schematic diagrams for me"

Scenarios where it should not be triggered:

Cover image requests → go to linyuebanzi-cover-hero skill
Documentary-type images such as product screenshots and error screenshots → these should be real screenshots and should not be generated by AI
Memes, avatars, single-image posters → use other tools
Pure code display → a code block is enough, no illustration needed
Xiaohongshu nine-grid → the style is completely different

Core process (four steps)

Step zero: Confirm the style

Before starting, first ask the user which style they want:

notebook (hand-drawn notebook style) Black ink as the main stroke + green for positive + red for emphasis, grid paper background. Suitable for engineering-oriented content, technical tutorials, and personal blogs.

"Just use the notebook hand-drawn style"

infographic (professional infographic style) Off-white background + dark brown-red headings, blue/orange dual-color contrast. Suitable for business presentations, comparative analysis, and brand content.

"Give me a professional infographic style one"

cozy-handdrawn (warm hand-drawn card style) Off-white paper-texture background + black hand-drawn lines + pastel rounded cards + light comic-style infographic layout + personal animated character to assist the narrative. Suitable for Chinese long-form technical articles, concept explanations, process breakdowns, comparative analysis, and WeChat official account tutorial illustrations.

"Use the warm hand-drawn card style"

tech-doodle (technical doodle style) Warm white cream background + fine-liner ink strokes + very light pastel coloring + minimalist stick-figure characters + highlighter quote highlights at the bottom. Suitable for technical blogs, knowledge management, architecture explanations, and personal brand content.

"Use the technical doodle style"

cartoon-infographic (cartoon infographic style) Clean white background + hand-drawn rounded containers + functional soft multi-color + optional LQ cartoon character to assist the narrative. Suitable for technical blogs, product documentation, knowledge management, and tutorial illustrations. Scenarios that need a little character automatically use edit mode to introduce a cartoon reference image; those that do not need one use pure generation.

"Use the cartoon infographic style"

whiteboard-sketch (whiteboard hand-drawn style) Warm cream background + black hand-drawn marker lines + soft pastel rounded containers + simple stick-figure characters + educational diagram layout. Like a schematic diagram casually drawn on a whiteboard by a skilled illustrator, friendly, clear, and warm. Suitable for technical tutorials, concept explanations, process descriptions, and architecture introductions.

"Use the whiteboard hand-drawn style"

If the user has no clear preference,You must first ask the user whether they want notebook, infographic, executive-tech, cozy-handdrawn, tech-doodle, cartoon-infographic, or whiteboard-sketch before continuing. Do not generate directly with the default style.

Step 1: Analyze the article and identify 4-5 positions that are "worth illustrating"

This is the most important step in the entire skill; it directly determines whether the article reads comfortably or not。

After reading the article, identify candidate positions according to the following four categories,and select a total of 4-5 based on each category's value density:

Type	When it counts as "worth illustrating"
🧠 Visualizing abstract concepts	The article is discussing a concept that is hard for readers to imagine out of thin air (such as "process isolation," "context pollution," "persona profile")
🔄 Process/loop/sequence	The article describes a multi-step process or loop (such as "task relay," "knowledge compounding," "pitfall loop")
⚖️ Comparison/classification	The article is making an A vs B comparison, or listing several categories (such as "implicit collaboration vs explicit collaboration," "three clone strategies")
🏗️ Architecture/component relationships	The article explains what parts a system consists of (such as "the composition of SOUL.md," "the role division of the three Agents")

Where not to illustrate:

Pure code sections (the code itself is already a visual element)
Places where the article already uses markdown tables, lists, or code blocks
Factual statements that can be explained clearly in one sentence
Emotional passages (complaints, punchlines, reflections—these rely on the rhythm of the text, and images actually interrupt it)

Distribution rhythm for 4-5 images:

Beginning (first 300 characters): do not place an image, let readers first enter the text
When the first core concept appears: image 1 (usually 🧠 visualizing abstract concepts)
Middle section: 2-3 images (interspersed with process/comparison/architecture diagrams)
Final punchline or summary paragraph: 1 image (optional, only add it if the article has an elevated summary)
After the ending hook: do not add images

After identifying them, record the selected positions using the structure below, and show it to the user before generating images:

我在文章里挑出了这 5 个建议插图的位置:

1. 【第 X 段,讲 profile 是什么】
   类型:🧠 抽象概念可视化
   图名:「Agent Profile & 真隔离」
   核心要表达:profile 是一个 AI 分身,由多个文件组成,靠进程隔离

2. 【第 Y 段,讲三个 Agent 的分工】
   类型:🏗️ 架构/组件关系
   图名:「三人小组的角色分工」
   核心要表达:林小管调度 / 林小探搜集 / 林小墨整理

...(共 4-5 个)

要按这个方案生图吗?或者你想调整某几处?

Wait for user confirmation before proceeding to step 2. This step cannot be skipped, because:

The user may want to preserve the "text purity" of certain positions and not add images
The user may want to add images at positions you did not select
This is a collaborative decision point; skipping it will make the user feel out of control

Step 2: Generate prompts according to the style template

Read references/prompt_template.md to get the skeleton template, and fill in the article material.Only write the chart content description, not the style prefix—the style prefix is automatically injected in step 3 by inject_style.py .

For each image, use the skeleton template and fill in the following information:

The image title(Chinese + English)
Main structure(whether it is "hub-and-spoke" / "process loop" / "left-right comparison" / "layered architecture")
Chinese label + annotation for each node
Emphasis points
Bottom summary text(give the image a punchline ending)

Style prefix(provided by references/styles/{style}.md,do not change the style itself) — The notebook style is a black-green-red grid paper baseline; the infographic style is an off-white background with blue-orange two-color infographic baseline; the executive-tech style is a deep indigo-purple primary color + card-based UI + magazine-style business tech aesthetic; the cozy-handdrawn style is an off-white paper-feel background + black hand-drawn lines + pastel rounded cards + a personal animated character assisting the narrative; the tech-doodle style is a warm white cream background + fine-liner ink strokes + very light pastel coloring + minimalist stick figures + a highlighter quote at the bottom; the cartoon-infographic style is a clean white background + hand-drawn rounded containers + functional soft multi-color + an optional LQ cartoon character.

Step 3: Inject the style and call the script to generate

3a. Inject the style prefix

Use scripts/inject_style.py to inject the user-selected style prefix into the prompt:

# 先写好不含风格前缀的 manifest JSON(diagrams.json),然后注入:
python3 scripts/inject_style.py \
    --style {用户选择的风格} \
    --manifest diagrams.json
# 输出: diagrams-styled.json(已注入风格前缀的完整 manifest)

Single-file injection is also supported (for re-running a specific image):

python3 scripts/inject_style.py \
    --style notebook \
    --prompt-file diagram-01.txt
# 输出: diagram-01-styled.txt

Note: the prompt in the manifest JSON should only contain the chart content description (do not write the style prefix); the style prefix is automatically injected by inject_style.py .inject_style.py automatically handles duplicate injection (removes the old prefix when switching styles and re-running).

3b. Call image-gen to generate

Use the script from the linyuebanzi-image-gen skill, passing in the manifest with the style injected:

python3 linyuebanzi-image-gen/scripts/generate.py \
    --manifest diagrams-styled.json \
    --output-dir ./diagrams/xxx \
    --parallel

The script will:

POST each image to the generation API and get a task_id
Poll the task until completion
Download the images locally
Save the prompt and metadata for each image

Parameters:

In the manifest, mode: generation(plain text image generation)
aspect_ratio: 16:9(script default)
resolution: 2K(script default)

Authentication: read from the environment variable MULERUN_API_KEY .

Manifest JSON format (before style injection)

After writing the chart content description in step 2, generate the following JSON (does not include the style prefix):

{
  "mode": "generation",
  "items": [
    {
      "id": "diagram-01-profile-isolation",
      "prompt": "{图表内容提示词,不含风格前缀}"
    },
    {
      "id": "diagram-02-role-division",
      "prompt": "{图表内容提示词,不含风格前缀}"
    },
    {
      "id": "diagram-03-flow-with-character",
      "prompt": "{图表内容提示词,不含风格前缀}",
      "needs_character": true
    }
  ]
}

Then use inject_style.py --manifest diagrams.json --style notebook to generate diagrams-styled.json, then call image-gen。

Generate the illustration placement list

After the script finishes running, based on the item ids in the results and the insertion position information recorded in step 1, manually generate a diagrams_manifest.md:

# 本次生成的插图 · 插入位置清单

## 插图 1: diagram-01-profile-isolation.png
**建议插入位置**:在"先聊聊 profile"这段后面

## 插图 2: diagram-02-role-division.png
**建议插入位置**:在"Step 1:建三个 Agent"小节开头

Artifact structure

diagrams/2026-04-21-hermes-multi-agent/
├── diagram-01-profile-isolation.png       # 第 1 张图
├── diagram-01-profile-isolation.txt       # 第 1 张图的提示词
├── diagram-02-role-division.png           # 第 2 张图
├── diagram-02-role-division.txt           # 第 2 张图的提示词
├── ...
└── diagrams_manifest.md                   # 插入位置清单

diagrams_manifest.md looks like this:

# 本次生成的插图 · 插入位置清单

## 插图 1: diagram-01-profile-isolation.png
**建议插入位置**:在"先聊聊 profile,这是整个多 Agent 的基础"小节下,"底层实现其实挺朴素"这段后面

## 插图 2: diagram-02-role-division.png
**建议插入位置**:在"Step 1:建三个 Agent,分工明确"小节开头,列出三个 Agent 名字后

...

Users just need to insert the images into the article according to this list.

Common pitfalls and countermeasures

Pitfall	Countermeasure
Garbled rendering of Chinese characters	Nano Banana 2 has better Chinese support than versions with lower Chinese character density, but it is still recommended to label key text bilingually in English + Chinese ; if all else fails, cover it up with PS afterward
The image is too flashy and loses the hand-drawn style	Repeatedly emphasize at the end of the prompt `clean minimal layout`、`hand-drawn sketch`、`no photographs`、`no 3D rendering`
Green and red used the wrong way round	Green = positive/correct process/growth, red = emphasis/warning/key point. Specify clearly in the prompt `Green for positive, red for emphasis only`
Inconsistent style between images	Use the same style declaration paragraph for all images, only change the main content
Images are disconnected from the article context	The "core message" of each image must be based on evidence from the original article text; do not improvise freely

References

Style templates:references/styles/notebook.md(hand-drawn notebook style) | references/styles/infographic.md(infographic style) | references/styles/executive-tech.md(modern high-end tech business style) | references/styles/cozy-handdrawn.md(warm hand-drawn card style) | references/styles/tech-doodle.md(technical stick-figure doodle style) | references/styles/cartoon-infographic.md(Cartoon infographic style) | references/styles/whiteboard-sketch.md(Whiteboard hand-drawn style)
Structured skeleton for hand-drawn image prompts:references/prompt_template.md
4 image type cases + 1 executive-tech complete style case:references/examples.md
General image generation script:linyuebanzi-image-gen/scripts/generate.py