ming-h Technical Diagram Generator
This skill helps you create professional, publication-quality technical diagrams for AI/ML technical articles using Mermaid, ASCII art, and structured formats.
When to Use This Skill
- Creating neural network architecture diagrams
- Visualizing data pipelines and ETL flows
- Documenting algorithm flows and decision trees
- Designing system architecture diagrams
- Explaining transformer attention mechanisms
- Showing training/inference workflows
- Creating comparison tables and matrices
Supported Diagram Types
1. Mermaid Diagrams
graph TB
Input[Input Data] --> Preprocess[Preprocessing]
Preprocess --> Model[Neural Network]
Model --> Output[Predictions]
2. ASCII Art Diagrams
+------------------+ +------------------+
| Input Layer | --> | Hidden Layer |
| (784 neurons) | | (256 neurons) |
+------------------+ +------------------+
|
v
+------------------+
| Output Layer |
| (10 neurons) |
+------------------+
3. Architecture Blocks
┌─────────────────────────────────────────┐
│ Transformer Block │
├─────────────────────────────────────────┤
│ ┌─────────────┐ ┌─────────────┐ │
│ │Multi-Head │ │ Feed │ │
│ │Attention │───>│ Forward │ │
│ └─────────────┘ └─────────────┘ │
│ │ │ │
│ └───────┬───────────┘ │
│ ▼ │
│ ┌─────────────┐ │
│ │Layer Norm │ │
│ └─────────────┘ │
└─────────────────────────────────────────┘
Instructions
For Neural Network Architectures
When asked to visualize a neural network:
## Network Architecture: [Name]
### Overview
[Brief description of the architecture]
### Architecture Diagram
```mermaid
graph TB
subgraph Input
X[Input Features<br/>shape: (batch, seq, dim)]
end
subgraph Encoder
E1[Embedding Layer]
E2[Positional Encoding]
E3[Multi-Head Attention]
E4[Layer Norm + FFN]
end
subgraph Output
O1[Linear Projection]
O2[Softmax]
Y[Predictions]
end
X --> E1 --> E2 --> E3 --> E4 --> O1 --> O2 --> Y
Layer Details
| Layer | Input Shape | Output Shape | Parameters |
|---|---|---|---|
| Embedding | (B, L) | (B, L, D) | V × D |
| Attention | (B, L, D) | (B, L, D) | 4D² |
| FFN | (B, L, D) | (B, L, D) | 2D × 4D |
Key Components
- Embedding Layer: Maps tokens to dense vectors
- Positional Encoding: Adds position information
- Multi-Head Attention: Computes attention weights
- Feed-Forward Network: Non-linear transformation
### For Data Pipelines
When asked to visualize a data pipeline:
```markdown
## Data Pipeline: [Name]
### Pipeline Flow
┌──────────────┐ ┌──────────────┐ ┌──────────────┐ │ Raw Data │───>│ Validation │───>│ Transform │ │ (S3/GCS) │ │ & Clean │ │ & Feature │ └──────────────┘ └──────────────┘ └──────────────┘ │ ▼ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ │ Serving │<───│ Training │<───│ Data │ │ Layer │ │ Pipeline │ │ Splitting │ └──────────────┘ └──────────────┘ └──────────────┘
### Processing Steps
| Stage | Operation | Tools | Output |
|-------|-----------|-------|--------|
| Ingestion | Read from source | Spark, Beam | Raw DataFrame |
| Validation | Schema check | Great Expectations | Cleaned DataFrame |
| Transform | Feature engineering | Pandas, PySpark | Feature Matrix |
| Split | Train/Val/Test | Scikit-learn | Split Data |
For Algorithm Flows
When asked to visualize an algorithm:
## Algorithm: [Name]
### Flowchart
```mermaid
flowchart TD
A[Start] --> B{Condition?}
B -->|Yes| C[Action A]
B -->|No| D[Action B]
C --> E[Process]
D --> E
E --> F{Converged?}
F -->|No| B
F -->|Yes| G[End]
Pseudocode
algorithm AlgorithmName(input):
# Step 1: Initialization
initialize parameters
# Step 2: Main Loop
while not converged:
# Step 2.1: Compute
result = compute(input)
# Step 2.2: Update
parameters = update(result)
# Step 2.3: Check
if convergence_criteria():
break
return output
Complexity Analysis
| Metric | Value |
|---|---|
| Time Complexity | O(n log n) |
| Space Complexity | O(n) |
| Best Case | O(n) |
| Worst Case | O(n²) |
### For Attention Mechanisms
When asked to explain attention:
```markdown
## Attention Mechanism
### Scaled Dot-Product Attention
Query (Q)
│
▼
┌─────────────────────────────┐ │ │ │ Q × K^T │ │ │ │ │ ▼ │ │ / √d_k (Scale) │ ←── Key (K) │ │ │ │ ▼ │ │ Softmax │ │ │ │ │ ▼ │ │ × V (Weighted Sum) │ ←── Value (V) │ │ │ │ ▼ │ │ Output │ │ │ └─────────────────────────────┘
### Multi-Head Attention
```mermaid
graph LR
Q[Query] --> H1[Head 1]
Q --> H2[Head 2]
Q --> H3[Head 3]
Q --> Hn[Head N]
K[Key] --> H1
K --> H2
K --> H3
K --> Hn
V[Value] --> H1
V --> H2
V --> H3
V --> Hn
H1 --> C[Concat]
H2 --> C
H3 --> C
Hn --> C
C --> L[Linear]
L --> O[Output]
Mathematical Formulation
$$\text{Attention}(Q, K, V) = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V$$
Where:
- $Q \in \mathbb{R}^{n \times d_k}$ (Query matrix)
- $K \in \mathbb{R}^{n \times d_k}$ (Key matrix)
- $V \in \mathbb{R}^{n \times d_v}$ (Value matrix)
- $d_k$ = dimension of keys
### For Training Workflows
When asked to show training process:
```markdown
## Training Workflow
### Training Loop
```mermaid
sequenceDiagram
participant D as DataLoader
participant M as Model
participant L as Loss Function
participant O as Optimizer
loop For each epoch
D->>M: Forward pass (batch)
M->>L: Compute loss
L->>O: Backward pass
O->>M: Update weights
end
Note over M: Model converged!
Training Pipeline
┌────────────────────────────────────────────────────────┐
│ Training Loop │
├────────────────────────────────────────────────────────┤
│ │
│ ┌─────────┐ ┌─────────┐ ┌─────────┐ │
│ │ Batch │──>│ Forward │──>│ Loss │ │
│ │ Data │ │ Pass │ │ Calc │ │
│ └─────────┘ └─────────┘ └─────────┘ │
│ │ │
│ ▼ │
│ ┌─────────┐ ┌─────────┐ ┌─────────┐ │
│ │ Update │<──│Gradient │<──│Backward │ │
│ │ Weights │ │ Calc │ │ Pass │ │
│ └─────────┘ └─────────┘ └─────────┘ │
│ │
└────────────────────────────────────────────────────────┘
Hyperparameters
| Parameter | Value | Description |
|---|---|---|
| Learning Rate | 1e-4 | Step size for gradient descent |
| Batch Size | 32 | Samples per gradient update |
| Epochs | 100 | Full dataset iterations |
| Optimizer | AdamW | Adaptive learning rate |
| Scheduler | Cosine | LR decay strategy |
### For System Architectures
When asked to show system design:
```markdown
## System Architecture: [Name]
### High-Level Architecture
```mermaid
graph TB
subgraph Client
UI[Web UI]
API[API Client]
end
subgraph Load Balancer
LB[Nginx/ALB]
end
subgraph Application
API1[API Server 1]
API2[API Server 2]
end
subgraph Services
Auth[Auth Service]
ML[ML Service]
Cache[Redis Cache]
end
subgraph Data
DB[(PostgreSQL)]
S3[(S3 Storage)]
end
UI --> LB
API --> LB
LB --> API1
LB --> API2
API1 --> Auth
API2 --> ML
ML --> Cache
Auth --> DB
ML --> S3
Component Details
| Component | Technology | Purpose |
|---|---|---|
| Frontend | React | User interface |
| API Gateway | FastAPI | Request routing |
| Auth | JWT | Authentication |
| ML Service | PyTorch | Model inference |
| Cache | Redis | Response caching |
| Database | PostgreSQL | Data persistence |
| Storage | S3 | Model artifacts |
## Best Practices
### Diagram Design Principles
1. **Clarity First**: Diagrams should be immediately understandable
2. **Consistent Styling**: Use consistent colors, shapes, and fonts
3. **Logical Flow**: Information should flow naturally (top-to-bottom or left-to-right)
4. **Appropriate Detail**: Show enough detail without overwhelming
5. **Clear Labels**: Every component should be labeled
### Color Coding
┌─────────────────────────────────────────┐ │ Color Coding Guide │ ├─────────────────────────────────────────┤ │ 🔵 Blue - Input/Output layers │ │ 🟢 Green - Processing/Transform │ │ 🟡 Yellow - Decision points │ │ 🔴 Red - Error/Warning states │ │ 🟣 Purple - Model components │ │ ⚪ Gray - Infrastructure │ └─────────────────────────────────────────┘
### Shape Conventions
┌─────────────┐ Rectangle: Process/Data │ Process │ └─────────────┘
◇ Diamond: Decision
/ \
/
◇───────◇
(Circle) Circle: Start/End
[/Parallel] Parallelogram: I/O
## Quick Reference
### Mermaid Syntax Cheat Sheet
graph TB # Top-bottom flowchart graph LR # Left-right flowchart flowchart TD # Flowchart with decisions sequenceDiagram # Sequence diagram classDiagram # UML class diagram stateDiagram # State machine pie # Pie chart gantt # Gantt chart
### Common Patterns
**Neural Network Layer**:
```mermaid
graph LR
Input[Input<br/>B×L×D] --> Linear[Linear<br/>D→H]
Linear --> Activation[ReLU/GELU]
Activation --> Output[Output<br/>B×L×H]
Attention Block:
graph TB
X[Input] --> QKV[Q, K, K Projection]
QKV --> Attn[Scaled Dot-Product]
Attn --> Proj[Output Projection]
Proj --> Add[Residual Add]
Add --> Out[Output]
Data Flow:
graph LR
A[Source] -->|Read| B[Transform]
B -->|Process| C[Validate]
C -->|Write| D[Sink]
Examples
Example 1: Transformer Architecture
## Transformer Architecture
### Complete Architecture
```mermaid
graph TB
subgraph Input
Tok[Tokenization]
Pos[Positional Encoding]
end
subgraph Encoder
E_MHA[Multi-Head Attention]
E_FF[Feed Forward]
E_N1[Layer Norm]
E_N2[Layer Norm]
end
subgraph Decoder
D_MHA1[Masked Attention]
D_MHA2[Cross Attention]
D_FF[Feed Forward]
D_N1[Layer Norm]
D_N2[Layer Norm]
D_N3[Layer Norm]
end
subgraph Output
Linear[Linear]
Softmax[Softmax]
end
Tok --> Pos --> E_MHA --> E_N1 --> E_FF --> E_N2
E_N2 --> D_MHA2
Pos --> D_MHA1 --> D_N1 --> D_MHA2 --> D_N2 --> D_FF --> D_N3 --> Linear --> Softmax
### Example 2: Diffusion Model
```markdown
## Diffusion Model Architecture
### Forward & Reverse Process
Forward Process (Add Noise)
───────────────────────────>
x₀ ──► x₁ ──► x₂ ──► ... ──► xₜ ──► ... ──► x_T 🖼️ 🌫️ 🌫️ 🌫️ 📺
<───────────────────────────
Reverse Process (Denoise)
x₀ ←── x₁ ←── x₂ ←── ... ←── xₜ ←── ... ←── x_T
🖼️ 🌫️ 🌫️ 🌫️ 📺
│ │ │
▼ ▼ ▼
ε̂₁ ε̂₂ ... ε̂ₜ
(Predicted Noise)
### U-Net Denoiser
```mermaid
graph TB
subgraph Encoder
E1[Conv 64]
E2[Conv 128]
E3[Conv 256]
E4[Conv 512]
end
subgraph Bottleneck
B[Attention Blocks]
end
subgraph Decoder
D4[UpConv 256]
D3[UpConv 128]
D2[UpConv 64]
D1[Output Conv]
end
x_t --> E1 --> E2 --> E3 --> E4 --> B
B --> D4 --> D3 --> D2 --> D1 --> ε̂
E1 -.-> D2
E2 -.-> D3
E3 -.-> D4
## Tips for AI/ML Diagrams
1. **Show Tensor Shapes**: Always include dimensions (B, L, D)
2. **Mark Trainable Parameters**: Use different colors for learnable vs fixed
3. **Highlight Data Flow**: Use arrows to show information flow
4. **Include Mathematical Notation**: Add equations where helpful
5. **Show Batch Dimension**: Always include B for batch
6. **Use Subgraphs**: Group related components together