---

name: presentation-generator description: "Generate high-quality technical HTML presentations (Reveal.js) and Markdown technical deep-dive articles from projects or papers. Covers architecture diagrams, code walkthroughs, math formulas, and speaker notes. Use when users say: 'make a presentation', 'create slides', 'generate HTML slides', 'technical presentation', 'paper presentation', 'Reveal.js slides', 'write a technical deep-dive article', 'create tech blog from code', 'generate architecture diagrams with matplotlib'. Triggers on requests to: (1) Convert a project/paper into an HTML slide deck, (2) Write a Markdown technical analysis article, (3) Generate matplotlib architecture diagrams for documentation."

Technical Presentation & Article Generator

Generate Reveal.js HTML presentations and Markdown technical deep-dive articles from complex technical projects or papers. All content in Chinese by default.

Core Principles

• Logic First: Establish overall architecture before diving into details. No fragmented info dumps.
• Visualization Driven: Use diagrams (Mermaid, architecture charts) over plain text whenever possible.
• Code-Theory Alignment: Every formula/diagram must map to concrete code implementation.
• Deep Dive via "What-Why-How-Code": For each core component, follow the 5-step pattern: Concept → Motivation → Mechanism → Code → Comparison.
• User Friendly: Leverage HTML/JS features for code highlighting, speaker notes, and interactivity.

Workflow: HTML Presentation Generation

Step 1: Analyze Source Material

Read the project/paper thoroughly. Collect:

• Model configs (yaml/json) for precise parameters
• Core model code for architecture details and tensor shapes
• Training code for loss design
• Inference code for inference flow
• README / paper for high-level overview

All technical details (parameter values, shapes, formulas) must come from source code, not guesswork.

Step 2: Plan Content Structure

Organize slides into these modules:

1. Cover & Context — Title, pain points, one-line innovation summary
2. High-Level Architecture — System overview, data flow, module interactions
3. Core Component Deep Dives (~50%+ of slides) — Detailed breakdown per module
4. Training & Inference Pipelines (AI model/algorithm only) — See Step 4.5 below
5. Performance & Optimization — Experiments, speedup strategies, latency
6. Summary & Future Work — Contributions, open problems

Step 3: Generate Single-File HTML

Produce a self-contained HTML file using Reveal.js via CDN. Must include:

• MathJax/KaTeX for LaTeX formulas
• highlight.js for code syntax highlighting
• Mermaid.js for diagrams (with manual rendering fix)
• Speaker notes (<aside class="notes">) on every slide — conversational style, explain "why" not "what"

Reveal.js Config (mandatory):

Reveal.initialize({
    width: 1920,
    height: 960,     // 2:1 aspect ratio for widescreen
    margin: 0.1,
    minScale: 0.2,
    maxScale: 1.5,
    center: true,
});

Centering CSS (mandatory):

.reveal .slides { text-align: center; }
.reveal .slides section {
    display: flex; flex-direction: column;
    justify-content: center; align-items: center;
    width: 100%; height: 100%;
    padding: 20px 40px; box-sizing: border-box;
}
.grid-2, .grid-3 { max-width: 1700px; width: 100%; }
.reveal .slides section > * { max-width: 100%; box-sizing: border-box; }

Step 4: Apply Deep Dive Pattern for Each Core Component

For every core module, follow this 5-step structure:

1. What: Definition, input/output
2. Why: Design motivation, what problem does it solve
3. How: Algorithm flowchart, core formulas (MathJax), shape flow (e.g., (B,C,T) → (B,2C,1))
4. Code: Key code snippet with line-level comments and highlighting
5. Comparison: Table comparing old vs new approach

Step 4.5: Training vs Inference Split (AI Model/Algorithm Projects)

When the project involves an AI model or algorithm, the training and inference pipelines MUST be presented as separate, clearly distinguished sections. Do NOT merge them into a single "model overview" slide.

Detection Criteria

If the source material contains ANY of the following, apply this step:

• Training scripts (train.py, trainer.py, fit(), loss functions)
• Inference scripts (infer.py, predict.py, generate())
• Distinct training-only components (data augmentation, loss design, learning rate schedule, gradient accumulation)
• Distinct inference-only components (beam search, sampling strategy, post-processing, quantization, TensorRT)
• Model behavior differences between training and inference (e.g., dropout, batch norm, teacher forcing vs autoregressive)

Required Slide Structure

Slide Group A: Training Pipeline (2-4 slides)

1. Training Architecture Overview — Training-specific data flow diagram showing: Dataset → Preprocessing → Model (train mode) → Loss → Optimizer → Update
2. Training Core Details — Loss function design (formulas + code), optimizer config, LR schedule, regularization strategies
3. Training Data Flow — Shape transformations specific to training (include batch dimension, label handling)
4. Training Tricks & Optimization (optional) — Mixed precision, gradient accumulation, distributed training, curriculum learning

Slide Group B: Inference Pipeline (2-4 slides)

1. Inference Architecture Overview — Inference-specific data flow diagram showing: Input → Preprocessing → Model (eval mode) → Post-processing → Output
2. Inference Core Details — Decoding strategy (greedy/beam/sampling), post-processing, confidence thresholds
3. Inference Data Flow — Shape transformations specific to inference (note differences from training: no labels, potentially different batch handling)
4. Inference Optimization (optional) — Quantization, pruning, caching (KV-cache), batching strategies, latency benchmarks

Slide Group C: Training vs Inference Comparison (1 slide)

Must include a comparison table:

Aspect	Training	Inference
Mode	model.train()	model.eval()
Data	Labeled dataset + augmentation	Raw input only
Dropout/BN	Active / running stats update	Disabled / frozen stats
Output	Loss value	Predictions
Batch Size	Large (throughput)	Small/1 (latency)
Key Metric	Training loss, validation accuracy	Latency, throughput, quality
Unique Components	Loss fn, optimizer, scheduler	Decoder, post-processor, cache

Visual Differentiation

• Use distinct color schemes: Training slides use 🔵 blue tones (#e3f2fd, #1976d2), Inference slides use 🟢 green tones (#e8f5e9, #388e3c)
• Use labeled section headers: "🏋️ Training Pipeline" and "🚀 Inference Pipeline"
• Mermaid/flowchart diagrams for training and inference should be separate diagrams, not a single combined one
• Highlight components that only exist in one phase (e.g., loss function is training-only; beam search is inference-only)

Step 5: Quality Checklist

Before delivering, verify:

• Every slide fits within viewport (no overflow/truncation)
• Font sizes appropriate (code not too small)
• Every "Why" is explained, not just "What"
• Core components have code correspondence
• Shape transformations clearly annotated
• Complex formulas have intuitive explanations
• All user-raised questions are addressed
• (AI model projects) Training and inference pipelines are presented separately with distinct visual styles
• (AI model projects) Training-only and inference-only components are clearly marked
• (AI model projects) A Training vs Inference comparison table is included

Workflow: Markdown Technical Article

Follow the article structure template in references/article-template.md.

Key principles:

• Table-driven: Use tables for parameter comparisons, model comparisons, shape references
• Code as documentation: Every core module needs code snippet + line comments
• Formula-code alignment: LaTeX symbols must match code variable names
• "Why" over "What": Explain design motivation for every decision

Workflow: Matplotlib Architecture Diagrams

See references/matplotlib-guide.md for the complete matplotlib diagram generation methodology.

Critical rules:

• FancyBboxPatch pad must be proportional: min(w, h) * 0.15
• Line height based on box height: h * 0.75 / n_lines
• Long-distance connections use polylines, not arcs
• Never use set_aspect('equal')
• macOS fonts: Arial Unicode MS > Heiti TC > STHeiti
• No emoji in matplotlib (not supported)

Common Pitfalls Quick Reference

Issue	Root Cause	Fix
Slide content overflow	Too much content per slide	Use scrollable containers or split slides
Content off-center	Missing Flexbox centering CSS	Apply mandatory centering CSS above
Mermaid not rendering on hidden slides	startOnLoad: true only renders visible	Set startOnLoad: false, manual render on ready + slidechanged
Vertical flowchart overflow	Too many nodes in HTML/CSS flowchart	Compress gap/padding/font-size, see references/revealjs-fixes.md
Matplotlib boxes distorted	pad in data coordinates, not pixels	Use proportional pad calculation
Chinese not showing in matplotlib	Wrong font	Use Arial Unicode MS / Heiti TC
Training/Inference merged into one slide	AI model specifics lost	Split into separate slide groups with distinct color themes (blue=train, green=infer)

Prompt Template

For generating presentations with maximum quality, see references/prompt-template.md for a proven C.R.I.S.P principle prompt.

Slide Template

See assets/slide-template.html for a starter HTML template with all required configs pre-set.