presenting-conference-talks

$npx mdskill add Orchestra-Research/AI-Research-SKILLs/presenting-conference-talks

Generates Beamer and PPTX slides with speaker notes from research papers for conference talks

  • Converts compiled research papers into presentation slides for ML and systems conferences
  • Uses Python-pptx library to create editable PPTX alongside LaTeX Beamer PDFs
  • Analyzes paper structure to extract key sections and generate speaker notes
  • Delivers polished slides with speaker notes and optional talk script for oral presentations

SKILL.md

.github/skills/presenting-conference-talksView on GitHub ↗
---
name: presenting-conference-talks
description: Generates conference presentation slides (Beamer LaTeX PDF and editable PPTX) from a compiled paper with speaker notes and talk script. Use when preparing oral talks, spotlight presentations, or invited talks for ML and systems conferences.
version: 1.0.0
author: Orchestra Research
license: MIT
tags: [Presenting Conference Talks, Beamer, PPTX, Slides, Speaker Notes, OSDI, SOSP, ASPLOS, NeurIPS, ICML]
dependencies: [python-pptx>=0.6.21]
---

# Presenting Conference Talks: From Paper to Slides

Generate conference presentation slides from a compiled research paper. Produces both **Beamer LaTeX PDF** (for polished typesetting) and **editable PPTX** (for last-minute adjustments), with speaker notes and an optional talk script.

## When to Use This Skill

| Scenario | Use This Skill | Use Other Skills Instead |
|----------|---------------|--------------------------|
| Preparing oral/spotlight/poster-talk slides | ✅ | |
| Generating Beamer PDF + PPTX from paper | ✅ | |
| Speaker notes and talk script | ✅ | |
| Writing the paper itself | | ml-paper-writing |
| Structuring a systems paper | | systems-paper-writing |
| Creating publication-quality plots | | academic-plotting |

**Attribution**: This skill's structure draws inspiration from the ARIS paper-slides skill (570 lines, supporting poster/spotlight/oral/invited with Beamer+PPTX). This is an independent implementation for the AI-Research-SKILLs ecosystem.

---

## Talk Types and Slide Counts

| Talk Type | Duration | Slides | Content Depth |
|-----------|----------|--------|---------------|
| poster-talk | 3–5 min | 5–8 | Problem + key result only |
| spotlight | 5–8 min | 8–12 | Problem + approach + key results |
| oral | 15–20 min | 15–22 | Full story with evaluation highlights |
| invited | 30–45 min | 25–40 | Deep dive with context and demos |

**Rule of thumb**: ~1 slide per minute for oral, ~1.5 slides per minute for spotlight.

---

## Slide Structure Templates

### Poster-Talk (5–8 slides)

```text
Slide 1: Title + Authors + Affiliation
Slide 2: Problem — Why this matters (1 motivating figure)
Slide 3: Key Insight — One-sentence thesis
Slide 4: Approach Overview — Architecture diagram
Slide 5: Main Result — Headline numbers (1 figure)
Slide 6: Takeaway + QR code to paper/code
```

### Spotlight (8–12 slides)

```text
Slide 1:  Title + Authors
Slide 2:  Problem Statement — Concrete, quantified
Slide 3:  Motivation — Why existing solutions fall short
Slide 4:  Key Insight — Thesis statement
Slide 5:  System Overview — Architecture diagram
Slide 6:  Design Highlight 1 — Core mechanism
Slide 7:  Design Highlight 2 — Key innovation
Slide 8:  Evaluation Setup — Baselines and workloads (brief)
Slide 9:  Main Results — Headline performance figure
Slide 10: Ablation / Breakdown — What contributes most
Slide 11: Summary + Contributions
Slide 12: Thank You + Links
```

### Oral (15–22 slides)

```text
Slide 1:  Title + Authors + Venue
Slide 2:  Outline (optional — "roadmap" slide)
Slide 3:  Problem Context — Domain importance
Slide 4:  Problem Statement — Specific challenge
Slide 5:  Motivation — Gaps in existing systems
Slide 6:  Key Insight — Thesis
Slide 7:  System Overview — Architecture diagram
Slide 8:  Design Component 1 — Detailed walkthrough
Slide 9:  Design Component 2 — Detailed walkthrough
Slide 10: Design Component 3 — Detailed walkthrough
Slide 11: Design Alternatives — Why not other approaches
Slide 12: Implementation — Key engineering highlights
Slide 13: Evaluation Setup — Testbed, baselines, metrics
Slide 14: End-to-End Results — Main performance
Slide 15: Result Deep Dive — Breakdown or per-workload
Slide 16: Ablation Study — Component contributions
Slide 17: Scalability — Scaling behavior
Slide 18: Demo Slide (systems talks) — Screenshot or recording
Slide 19: Related Work — Positioning (brief)
Slide 20: Summary — Contributions restated
Slide 21: Future Work — Open questions
Slide 22: Thank You + Paper Link + QR Code
```

### Invited Talk (25–40 slides)
Extends the oral structure with:
- Additional context slides (field overview, historical progression)
- Multiple demo/walkthrough slides
- Deeper evaluation analysis
- Broader implications and future directions
- Q&A preparation slides (hidden, for backup)

---

## Systems Talk Specifics

Systems conference talks have unique requirements compared to ML talks:

### Demo Slide
- Include a **live demo** or **pre-recorded screencast** of the system in action
- Always have a **recorded backup** — live demos fail at the worst times
- Show the system under realistic load, not toy examples

### Architecture Walkthrough
- Animate the architecture diagram: highlight components as you explain them
- Use Beamer `\only<N>` or `\onslide<N>` for progressive reveal
- Walk through a **concrete request** end-to-end through the system

### Evaluation Highlights
- Select 2–3 strongest figures from the paper
- Annotate figures on slides (arrows, circles highlighting key points)
- State the takeaway **before** showing the figure ("Our system is 2x faster — here's the data")

---

## Speaker Notes Guidelines

### Structure per Slide
```text
[Timing: X minutes]
[Key point to convey]
[Transition sentence to next slide]
```

### Mike Dahlin's Layered Approach
Apply "Say what you're going to say, say it, then say what you said" at three levels:

1. **Talk level**: Outline slide → body → summary slide
2. **Section level**: Section heading → content slides → section takeaway
3. **Slide level**: Headline statement → supporting evidence → transition

### Timing Guidelines
- Poster-talk: 30–60 sec per slide
- Spotlight: 30–45 sec per slide
- Oral: 45–90 sec per slide
- Invited: 60–120 sec per slide

---

## Output Formats

### Beamer LaTeX → PDF

Advantages: Professional typesetting, math support, version control friendly.

```latex
\documentclass[aspectratio=169]{beamer}
\usetheme{metropolis}  % Clean, modern theme
\usepackage{appendixnumberbeamer}

\title{Your Paper Title}
\subtitle{Venue Year}
\author{Author 1 \and Author 2}
\institute{Institution}
\date{}

\begin{document}
\maketitle

\begin{frame}{Problem}
  \begin{itemize}
    \item Key problem statement
    \item Concrete motivation with numbers
  \end{itemize}
  \note{Speaker note: Start with the big picture...}
\end{frame}

% ... more frames ...
\end{document}
```

### python-pptx → Editable PPTX

Advantages: Easy last-minute edits, corporate template compatibility, animations.

```python
from pptx import Presentation
from pptx.util import Inches, Pt
from pptx.enum.text import PP_ALIGN

prs = Presentation()
prs.slide_width = Inches(13.333)  # 16:9
prs.slide_height = Inches(7.5)

# Title slide
slide = prs.slides.add_slide(prs.slide_layouts[0])
slide.shapes.title.text = "Your Paper Title"
slide.placeholders[1].text = "Author 1, Author 2\nVenue Year"

# Content slide
slide = prs.slides.add_slide(prs.slide_layouts[1])
slide.shapes.title.text = "Problem Statement"
body = slide.placeholders[1]
body.text = "Key point 1\nKey point 2"

# Add speaker notes
notes_slide = slide.notes_slide
notes_slide.notes_text_frame.text = "Speaker note: explain the motivation..."

prs.save("talk.pptx")
```

---

## Color Scheme Suggestions

> These are aesthetic suggestions, not official venue requirements. Adjust freely.

| Venue Type | Primary | Accent | Background |
|-----------|---------|--------|------------|
| USENIX (OSDI/NSDI) | Dark Blue (#003366) | Red (#CC0000) | White |
| ACM (SOSP/ASPLOS) | ACM Blue (#0071BC) | Dark Gray (#333333) | White |
| NeurIPS | Purple (#7B2D8E) | Gold (#F0AD00) | White |
| ICML | Teal (#008080) | Orange (#FF6600) | White |
| Generic | Dark Gray (#333333) | Blue (#0066CC) | White |

---

## Workflow

### Step 1: Content Extraction
```text
- Read the compiled paper (PDF or LaTeX source)
- Identify: thesis, contributions, architecture figure, key eval figures
- Note the talk type and duration
```

### Step 2: Outline Generation
```text
- Select the appropriate slide structure template (above)
- Map paper sections to slide groups
- Allocate time per slide group
```

### Step 3: Slide-by-Slide Generation
```text
- Generate Beamer source slide by slide
- Add speaker notes per slide
- Include figures from paper (copy to slides/ directory)
- Generate python-pptx script for PPTX version
```

### Step 4: Review and Polish
```text
- Check total slide count matches talk duration
- Verify all figures are readable at presentation resolution
- Run Beamer compilation: latexmk -pdf slides.tex
- Run PPTX generation: python3 generate_slides.py
- Review speaker notes for timing and transitions
```

### Quick Checklist
- [ ] Slide count appropriate for talk type/duration
- [ ] Title slide has correct authors, affiliations, venue
- [ ] Architecture diagram included and clearly labeled
- [ ] Key eval figures annotated with takeaways
- [ ] Speaker notes include timing markers
- [ ] Transitions between sections are smooth
- [ ] Demo slide has recorded backup
- [ ] Thank-you slide includes paper link / QR code
- [ ] Font sizes ≥ 24pt for readability from back of room
- [ ] Consistent color scheme throughout

---

## Common Issues and Solutions

| Issue | Solution |
|-------|----------|
| Too many slides for time limit | Cut details, keep one figure per point |
| Slides feel like paper paragraphs | Use bullet points (≤ 6 per slide), let figures tell the story |
| Audience lost during design section | Add architecture walkthrough with progressive reveal |
| Evaluation slides overwhelming | Show 2–3 strongest figures, put rest in backup slides |
| Speaker notes too long | Target 3–4 sentences per slide, focus on transitions |
| Beamer compilation fails | Check figure paths, use `\graphicspath{{figures/}}` |
| PPTX looks different from Beamer | Adjust python-pptx font sizes and margins manually |

---

## References

- [references/slide-templates.md](references/slide-templates.md) — Complete Beamer template code and python-pptx generation script
- Mike Dahlin, "Giving a Conference Talk" — https://www.cs.utexas.edu/~dahlin/professional/goodTalk.pdf

More from Orchestra-Research/AI-Research-SKILLs

SkillDescription
academic-plottingGenerates publication-quality figures for ML papers from research context. Given a paper section or description, extracts system components and relationships to generate architecture diagrams via Gemini. Given experiment results or data, auto-selects chart type and generates data-driven figures via matplotlib/seaborn. Use when creating any figure for a conference paper.
ara-compilerCompiles any research input — PDF papers, GitHub repositories, experiment logs, code directories, or raw notes — into a complete Agent-Native Research Artifact (ARA) with cognitive layer (claims, concepts, heuristics), physical layer (configs, code stubs), exploration graph, and grounded evidence. Use when ingesting a paper or codebase into a structured, machine-executable knowledge package, building an ARA from scratch, or converting research outputs into a falsifiable, agent-traversable form.
ara-research-managerRecords research provenance as a post-task epilogue, scanning conversation history at the end of a coding or research session to extract decisions, experiments, dead ends, claims, heuristics, and pivots, and writing them into the ara/ directory with user-vs-AI provenance tags. Use as a session epilogue — never during execution — to maintain a faithful, auditable trace of how a research project actually evolved.
ara-rigor-reviewerPerforms ARA Seal Level 2 semantic epistemic review on Agent-Native Research Artifacts, scoring six dimensions (evidence relevance, falsifiability, scope calibration, argument coherence, exploration integrity, methodological rigor) and producing a constructive, severity-ranked report with a Strong Accept-to-Reject recommendation. Use after Level 1 structural validation passes, when an ARA needs an objective epistemic critique before publication or release.
autogpt-agentsAutonomous AI agent platform for building and deploying continuous agents. Use when creating visual workflow agents, deploying persistent autonomous agents, or building complex multi-step AI automation systems.
awq-quantizationActivation-aware weight quantization for 4-bit LLM compression with 3x speedup and minimal accuracy loss. Use when deploying large models (7B-70B) on limited GPU memory, when you need faster inference than GPTQ with better accuracy preservation, or for instruction-tuned and multimodal models. MLSys 2024 Best Paper Award winner.
blip-2-vision-languageVision-language pre-training framework bridging frozen image encoders and LLMs. Use when you need image captioning, visual question answering, image-text retrieval, or multimodal chat with state-of-the-art zero-shot performance.
brainstorming-research-ideasGuides researchers through structured ideation frameworks to discover high-impact research directions. Use when exploring new problem spaces, pivoting between projects, or seeking novel angles on existing work.
constitutional-aiAnthropic's method for training harmless AI through self-improvement. Two-phase approach - supervised learning with self-critique/revision, then RLAIF (RL from AI Feedback). Use for safety alignment, reducing harmful outputs without human labels. Powers Claude's safety system.
creative-thinking-for-researchApplies cognitive science frameworks for creative thinking to CS and AI research ideation. Use when seeking genuinely novel research directions by leveraging combinatorial creativity, analogical reasoning, constraint manipulation, and other empirically grounded creative strategies.