🧠 Code Analysis — 代码逻辑图自动生成器

A Claude Code Skill that automatically scans project directories and generates high-resolution bilingual (EN/CN) code logic diagrams — tailored for DL/ML model architectures and multi-language system design.

---

✨ Features

Feature	Description
🔍 Smart Scan	Auto-detect model definitions (PyTorch/TF/ONNX/Flax), training pipelines, and service code
🏗️ DL Model Architecture	Paper-style model diagrams tracing forward()/create_model() execution path
📊 Code Logic Diagrams	Call graphs, sequence diagrams, data-flow charts, class diagrams
🌐 Bilingual Labels	All nodes/edges in English + 中文 for cross-team readability
🔠 Large Font Rendering	Mermaid config-based 24px font sizing — text never overflows node boundaries
🖼️ High-res Output	4× scale, 3840×2880 canvas — sharp even at poster-size prints
🎨 Professional Styling	Custom classDef color palette, grouped subgraphs, labeled edges
🚀 Zero Config	Auto-installs mermaid-cli via npx if not present

---

📂 Repository Structure

Code_Analysis/
├── README.md                          ← You are here
├── code-diagram-generator/            ← Claude Code Skill
│   └── SKILL.md                       ← Skill definition (install this)
└── diagrams/                          ← Example outputs (faster-whisper project)
    ├── mermaid-config.json            ← Font config (24px, spacing)
    ├── large-font.css                 ← CSS supplementary protection
    ├── model/
    │   ├── whisper_architecture.mmd   ← Whisper model E2E architecture
    │   ├── whisper_architecture.svg
    │   └── whisper_architecture.png   (4× scale)
    ├── logic/
    │   ├── transcription_call_graph.* ← Full module call graph
    │   ├── transcription_sequence.*   ← 7-phase transcription timeline
    │   ├── batched_vs_sequential.*    ← Two inference mode comparison
    │   └── project_structure.*        ← Package structure & deps
    └── pipeline/
        ├── inference_pipeline.mmd     ← End-to-end inference data flow
        ├── inference_pipeline.svg
        └── inference_pipeline.png     (4× scale)

---

📸 Example Outputs

Whisper Model Architecture 模型架构

The complete faster-whisper inference pipeline: Audio Input → PyAV Decode → Silero VAD → STFT/Mel Feature Extraction → CTranslate2 Encoder (Conv1D + Transformer) → Decoder (Token Embedding + Cross-Attention) → Beam Search with Temperature Fallback → Word-level Alignment via DTW.

---

Transcription Call Graph 转录调用图

Full call relationship across all modules: WhisperModel.transcribe() and BatchedInferencePipeline.transcribe() entry points, spanning audio.py, feature_extractor.py, tokenizer.py, vad.py, and the CTranslate2 backend.

---

Transcription Sequence 转录时序图

Complete 7-phase lifecycle of a single transcribe() call: Audio Decode → VAD Filter → Feature Extraction → Language Detection → Sliding Window Generation (with fallback) → Word Timestamps → Timestamp Restoration.

---

Sequential vs Batched Mode 顺序模式 vs 批处理模式

Architecture comparison of the two inference modes: WhisperModel.transcribe() (sliding window, condition on previous text) vs BatchedInferencePipeline (VAD-chunked, parallel batch GPU inference).

---

Inference Pipeline 推理流水线

End-to-end data flow from audio input (file/stream/array) through decode, VAD, Mel spectrogram, CTranslate2 encoder/decoder, to segment/word output.

---

Project Structure 工程目录结构

Package layout and internal module dependencies of faster-whisper: transcribe.py (core, ~1900 lines) depending on feature_extractor.py, audio.py, tokenizer.py, vad.py, and utils.py.

---

🚀 How to Use (as a Claude Code Skill)

Install

1. Copy code-diagram-generator/SKILL.md to ~/.claude/skills/code-diagram-generator/SKILL.md
2. Or point Claude Code to this repo path

Invoke

Just tell Claude:

帮我分析 D:\my-project 的代码结构，生成逻辑图

Or in English:

Generate architecture diagrams for ./my-ml-project

Requirements

• Node.js ≥ 16 (for mermaid-cli)
• mermaid-cli auto-installs via npx

---

🔧 Rendering (Manual)

If you want to render .mmd files yourself:

# All-in-one command
npx @mermaid-js/mermaid-cli \
  -i input.mmd \
  -o output.png \
  -b white \
  -c mermaid-config.json \
  --width 3840 --height 2880 --scale 4 --pdfFit

Key file: diagrams/mermaid-config.json — this sets fontSize: "24px" at the layout engine level, so node containers are sized correctly for large bilingual text.

---

🎨 Design Decisions

Decision	Reasoning
Mermaid (not Graphviz/D2)	Text-based, version-controllable, GitHub-native rendering
Config-based font sizing (not CSS)	CSS can't change baked-in foreignObject dimensions — config makes layout engine compute correct sizes
--scale 4	4× oversampling produces print-quality PNGs (1-3 MB each)
Bilingual labels	Teams in China often mix EN/CN; single-language diagrams always leave someone out
classDef color palette	Mermaid defaults are pastel and hard to distinguish; professional colors improve scannability

---

📋 Supported Languages & Frameworks

Language	Frameworks Detected
Python	PyTorch, TensorFlow, Keras, JAX/Flax, PaddlePaddle, ONNX Runtime, Flask
Go	net/http, TF C API, ONNX Runtime Go, Redis
Rust	Actix-web, TF C FFI, ort crate, Redis
JavaScript/TypeScript	Node.js patterns, Express
General	Any project with function calls, classes, or data pipelines

---

License

MIT