paper-slides

name	paper-slides
description	Generate conference presentation slides (beamer LaTeX → PDF + editable PPTX) from a compiled paper, with speaker notes and full talk script. Use when user says "做PPT", "做幻灯片", "make slides", "conference talk", "presentation slides", "生成slides", "写演讲稿", or wants beamer slides for a conference talk.
argument-hint	[paper-directory-or-talk-length] [— style-ref: <source>]
allowed-tools	Bash(*), Read, Write, Edit, Grep, Glob, Agent, mcp__codex__codex, mcp__codex__codex-reply

Paper Slides: From Paper to Conference Talk

Generate conference presentation slides from: $ARGUMENTS

Context

This skill runs after Workflow 3 (/paper-writing). It takes a compiled paper and generates a presentation slide deck for conference oral talks, spotlight presentations, or poster lightning talks.

Unlike posters (single page, visual-first), slides tell a temporal story: each slide builds on the previous one, with progressive revelation of the research narrative. A good talk makes the audience understand why this matters before showing what was done.

Constants

• VENUE = NeurIPS — Target venue, determines color scheme. Supported: NeurIPS, ICML, ICLR, AAAI, ACL, EMNLP, CVPR, ECCV, GENERIC. Override via argument.
• TALK_TYPE = spotlight — Talk format. Options: oral (15-20 min), spotlight (5-8 min), poster-talk (3-5 min), invited (30-45 min). Determines slide count and content depth.
• TALK_MINUTES = 15 — Talk duration in minutes. Auto-adjusts slide count (~1 slide/minute for oral, ~1.5 slides/minute for spotlight). Override explicitly if needed.
• ASPECT_RATIO = 16:9 — Slide aspect ratio. Options: 16:9 (default, modern projectors), 4:3 (legacy).
• SPEAKER_NOTES = true — Generate \note{} blocks in beamer and corresponding PPTX notes. Set false for clean slides without notes.
• PAPER_DIR = paper/ — Directory containing the compiled paper.
• OUTPUT_DIR = slides/ — Output directory for all slide files.
• REVIEWER_MODEL = gpt-5.4 — Model used via Codex MCP for slide review.
• AUTO_PROCEED = false — At each checkpoint, always wait for explicit user confirmation.
• COMPILER = latexmk — LaTeX build tool.
• ENGINE = pdflatex — LaTeX engine. Use xelatex for CJK text.

💡 Override: /paper-slides "paper/" — talk_type: oral, venue: ICML, minutes: 20, aspect: 4:3

Optional: Style reference (— style-ref: <source>, opt-in)

Lets the user steer the talk's structural rhythm (story beats, theorem density, figure density inherited from the source paper) toward a reference paper. Default OFF — when the user does not pass — style-ref, do nothing differently from before.

Only when — style-ref: <source> appears in $ARGUMENTS, run the helper FIRST:

if [ ! -f tools/extract_paper_style.py ]; then
  echo "error: tools/extract_paper_style.py not found — re-run 'bash tools/install_aris.sh' to refresh the '.aris/tools' symlink (added in #174), or copy the helper manually from the ARIS repo" >&2
  exit 1
fi
CACHE=$(python3 tools/extract_paper_style.py --source "<source>")
case $? in
  0) ;;                                       # use $CACHE/style_profile.md as structural guidance
  2) echo "warning: style-ref skipped (missing optional dep)" >&2 ;;
  3) echo "error: --style-ref source failed; aborting slides" >&2 ; exit 1 ;;
  *) echo "error: helper failed unexpectedly; aborting slides" >&2 ; exit 1 ;;
esac

Sources accepted: local TeX dir / file, local PDF, arXiv id, http(s) URL. Overleaf URLs/IDs are rejected — clone via /overleaf-sync setup <id> first and pass the local clone path.

Strict rules (full contract in tools/extract_paper_style.py docstring):

• Use style_profile.md to align section-budget tendency and theorem-environment density. Talk-type slide count above still takes precedence.
• Never copy speaker-note prose, slide titles, or examples from anything reachable through the cache. The talk content is from the user's paper, not the reference.
• Never pass — style-ref (or the cache contents) to the GPT-5.4 reviewer sub-agent — the reviewer must judge the talk's clarity on its own merits.

Talk Type → Slide Count

Talk Type	Duration	Slides	Content Depth
poster-talk	3-5 min	5-8	Problem + 1 method slide + 1 result + conclusion
spotlight	5-8 min	8-12	Problem + 2 method + 2 results + conclusion
oral	15-20 min	15-22	Full story with motivation, method detail, experiments, analysis
invited	30-45 min	25-40	Comprehensive: background, related work, deep method, extensive results, discussion

Venue Color Schemes

Same as /paper-poster:

Venue	Primary	Accent	Background	Text
NeurIPS	#8B5CF6	#2563EB	#FFFFFF	#1E1E1E
ICML	#DC2626	#1D4ED8	#FFFFFF	#1E1E1E
ICLR	#059669	#0284C7	#FFFFFF	#1E1E1E
CVPR	#2563EB	#7C3AED	#FFFFFF	#1E1E1E
GENERIC	#334155	#2563EB	#FFFFFF	#1E1E1E

State Persistence (Compact Recovery)

Persist state to slides/SLIDES_STATE.json after each phase:

{
  "phase": 3,
  "venue": "NeurIPS",
  "talk_type": "spotlight",
  "slide_count": 10,
  "codex_thread_id": "019cfcf4-...",
  "status": "in_progress",
  "timestamp": "2026-03-18T15:00:00"
}

On startup: if SLIDES_STATE.json exists with "status": "in_progress" and within 24h → resume. Otherwise → fresh start.

Workflow

Phase 0: Input Validation & Setup

1. Check prerequisites:

   which pdflatex && which latexmk
   

2. Verify paper exists:

   ls $PAPER_DIR/main.tex || ls $PAPER_DIR/main.pdf
   ls $PAPER_DIR/sections/*.tex
   ls $PAPER_DIR/figures/
   

3. Backup existing slides: if slides/ exists, copy to slides-backup-{timestamp}/

4. Create output directory: mkdir -p slides/figures

5. Detect CJK: if paper contains Chinese/Japanese/Korean, set ENGINE to xelatex

6. Determine slide count: from TALK_TYPE and TALK_MINUTES using the table above

7. Check for resume: read slides/SLIDES_STATE.json if it exists

State: Write SLIDES_STATE.json with phase: 0.

Phase 1: Content Extraction & Slide Outline

Read paper/sections/*.tex and build a slide-by-slide outline.

Slide template by talk type:

Oral (15-22 slides)

Slide	Purpose	Content Source	Figure?
1	Title	Paper metadata	No
2	Outline	Section headers	No
3-4	Motivation & Problem	Introduction	Optional
5	Key Insight	Introduction (contribution)	No
6-9	Method	Method section	Yes (hero figure)
10-14	Results	Experiments	Yes (per slide)
15-16	Analysis / Ablations	Experiments	Yes
17	Limitations	Conclusion	No
18	Conclusion / Takeaway	Conclusion	No
19	Thank You + QR	—	QR code

Spotlight (8-12 slides)

Slide	Purpose	Content Source	Figure?
1	Title	Paper metadata	No
2-3	Problem + Why It Matters	Introduction	Optional
4	Key Insight	Contribution	No
5-6	Method	Method (condensed)	Yes (hero)
7-9	Results	Key results only	Yes
10	Takeaway	Conclusion	No
11	Thank You + QR	—	QR code

Poster-talk (5-8 slides)

Slide	Purpose	Content Source	Figure?
1	Title	Paper metadata	No
2	Problem	Introduction (1 slide)	No
3	Method	Method (1 slide)	Yes
4-5	Results	Key result only	Yes
6	Takeaway + QR	Conclusion	QR

For each slide, specify:

• Title (max 8 words)
• 3-5 bullet points (max 8 words each)
• Figure reference (if any) from paper/figures/
• Speaker note (2-3 sentences of what to say)
• Time allocation (in seconds)

Output: slides/SLIDE_OUTLINE.md

🚦 Checkpoint:

📊 Slide outline ready:
- Talk type: [TALK_TYPE] ([TALK_MINUTES] min)
- Slide count: [N] slides
- Figures used: [N] from paper/figures/
- Time budget: [breakdown]

Slide-by-slide outline:
1. [Title slide]
2. [Motivation — 1.5 min]
3. [Problem statement — 1 min]
...

Proceed to drafting? Or adjust the outline?

⛔ STOP HERE and wait for user response. This is the most critical checkpoint — the outline determines the entire talk flow.

Options:

• "go" → proceed to Phase 2
• adjustments (e.g., "merge slides 3-4", "add a demo slide", "cut the ablation") → revise
• "stop" → save to slides/SLIDE_OUTLINE.md

State: Write SLIDES_STATE.json with phase: 1.

Phase 2: Slide-by-Slide Content Drafting

For each slide in the outline, draft the actual content.

Presentation rules (enforced strictly):

Rule	Rationale
One message per slide	If a slide has two ideas, split it
Max 6 lines per slide	More than 6 lines = wall of text
Max 8 words per line	Audience reads, not listens, if text is long
Sentence fragments, not sentences	"Improves F1 by 3.2%" not "Our method improves the F1 score by 3.2 percentage points"
Figure slides: figure ≥60% area	The figure IS the content; bullets are annotations
Bold key numbers	"Achieves 94.3% accuracy"
Progressive disclosure	Use \pause or \onslide for complex slides
No Related Work slide	Unless invited talk (30+ min)

For each slide, produce:

1. \frametitle{}
2. Content (itemize or figure + caption)
3. \note{} with speaker text (if SPEAKER_NOTES=true)

Phase 3: Generate Slides LaTeX

Create slides/main.tex using beamer.

Template structure:

\documentclass[aspectratio=169]{beamer}

% Venue theme
\usepackage{xcolor}
\definecolor{primary}{HTML}{VENUE_PRIMARY}
\definecolor{accent}{HTML}{VENUE_ACCENT}

% Clean theme
\usetheme{default}
\usecolortheme{default}
\setbeamercolor{frametitle}{fg=primary}
\setbeamercolor{title}{fg=primary}
\setbeamercolor{structure}{fg=accent}
\setbeamercolor{itemize item}{fg=primary}
\setbeamercolor{itemize subitem}{fg=accent}
\setbeamertemplate{navigation symbols}{}
\setbeamertemplate{footline}{
  \hfill\insertframenumber/\inserttotalframenumber\hspace{2mm}\vspace{2mm}
}

% Packages
\usepackage{graphicx,amsmath,booktabs}
\graphicspath{{figures/}}

% Speaker notes (if enabled)
% \setbeameroption{show notes on second screen=right}

% Metadata
\title{PAPER TITLE}
\author{Author 1 \and Author 2}
\institute{Affiliation}
\date{VENUE YEAR}

\begin{document}

\begin{frame}
\titlepage
\end{frame}

% Content slides follow...

\begin{frame}{Motivation}
\begin{itemize}
  \item Bullet point 1
  \item Bullet point 2
  \item \textbf{Key insight in bold}
\end{itemize}
\note{Speaker note: explain the motivation...}
\end{frame}

% Figure slide example
\begin{frame}{Method Overview}
\centering
\includegraphics[width=0.85\textwidth]{method_overview.pdf}
\vspace{0.5em}
\begin{itemize}
  \item Key annotation about the figure
\end{itemize}
\note{Walk through the figure left to right...}
\end{frame}

% ... more slides ...

\begin{frame}{Thank You}
\centering
{\Large Questions?}\\[2em]
Paper: [URL or QR placeholder]\\
Code: [URL or QR placeholder]
\end{frame}

\end{document}

Symlink figures:

ln -sf ../paper/figures/*.pdf slides/figures/ 2>/dev/null
ln -sf ../paper/figures/*.png slides/figures/ 2>/dev/null

Key formatting rules:

• Title font: ≥28pt, venue primary color
• Body font: ≥20pt
• Footnotes: ≥14pt
• No navigation symbols
• Frame numbers in bottom-right
• Clean white background (no gradients, no decorative elements)

Phase 4: Compile Slides

cd slides && latexmk -$ENGINE -interaction=nonstopmode main.tex

Error handling loop (max 3 attempts):

1. Parse error log
2. Fix: missing package, undefined command, file not found, overfull boxes
3. Recompile

Verification:

# Check slide count matches outline
pdfinfo slides/main.pdf | grep Pages

If page count differs significantly from outline (>2 slides off), investigate.

State: Write SLIDES_STATE.json with phase: 4.

Phase 5: Codex MCP Review

Send the slide outline + selected LaTeX frames to GPT-5.4 xhigh:

mcp__codex__codex:
  config: {"model_reasoning_effort": "xhigh"}
  prompt: |
    Review this [TALK_TYPE] presentation ([TALK_MINUTES] min) for [VENUE].

    Evaluate using these criteria (score 1-5 each):

    1. **Story arc** — Does the talk build a compelling narrative? (Problem → insight → method → evidence → takeaway)
    2. **Slide density** — Any slides with too much text? (Max 6 lines, 8 words/line)
    3. **Time budget** — Is [N] slides realistic for [TALK_MINUTES] minutes?
    4. **Figure visibility** — Will figures be readable on a projector?
    5. **Opening hook** — Do slides 2-3 grab attention? (Not "In this paper, we...")
    6. **Takeaway** — Is the final message clear and memorable?
    7. **Progressive build** — Are complex ideas revealed gradually?

    Slide outline:
    [PASTE SLIDE_OUTLINE.md]

    Selected frames (LaTeX):
    [PASTE KEY FRAMES]

    Provide:
    - Score for each criterion
    - Top 3 actionable fixes
    - Overall: Ready to present? (Yes / Needs revision / Major issues)

Apply fixes. Recompile if LaTeX was changed.

⚠️ If mcp__codex__codex is not available (no OpenAI API key), skip external review and proceed to Phase 6. Note the skip in SLIDES_STATE.json.

Save review to slides/SLIDES_REVIEW.md.

State: Write SLIDES_STATE.json with phase: 5.

Phase 6: Speaker Notes

For each slide, ensure a \note{} block exists with:

1. What to say (2-3 complete sentences, conversational tone)
2. Timing hint (e.g., "spend 1 minute here", "quick — 20 seconds")
3. Transition phrase to the next slide (e.g., "So how do we actually implement this? Let me show you...")

Also generate slides/speaker_notes.md as a standalone backup:

# Speaker Notes

## Slide 1: Title
[No speaking — wait for introduction]

## Slide 2: Motivation
"Thank you. So let me start with the problem we're trying to solve..."
[Time: 1.5 min]

## Slide 3: Problem Statement
"Specifically, the challenge is..."
→ Transition: "To address this, our key insight is..."
[Time: 1 min]

...

State: Write SLIDES_STATE.json with phase: 6.

Phase 7: PowerPoint Export

Generate an editable PPTX using python-pptx:

python3 -c "import pptx" 2>/dev/null || pip install python-pptx

Write slides/generate_pptx.py that:

1. Creates a PPTX with correct aspect ratio (16:9 → 13.33" x 7.5"; 4:3 → 10" x 7.5")
2. For each beamer frame:
   • Creates a slide with matching layout
   • Title in venue primary color, bold
   • Bullet points with venue accent color markers
   • Figures embedded as images (from slides/figures/)
   • Speaker notes transferred to PPTX notes field
3. Title slide with special formatting (centered, larger title)
4. Thank You slide with centered text
5. Applies venue color scheme throughout

cd slides && python3 generate_pptx.py
# Output: slides/presentation.pptx

⚠️ If python-pptx is not installed, skip with a note: "Install pip install python-pptx to enable PowerPoint export."

State: Write SLIDES_STATE.json with phase: 7.

Phase 8: Full Talk Script

Generate slides/TALK_SCRIPT.md — a complete, word-for-word script for the talk.

This is different from speaker notes (brief reminders). The talk script is a full manuscript that can be read aloud or used for practice.

# Talk Script: [Paper Title]

**Venue**: [VENUE] [YEAR]
**Talk type**: [TALK_TYPE] ([TALK_MINUTES] min)
**Total slides**: [N]

---

## Slide 1: Title [0:00 - 0:15]

*[Wait for chair introduction]*

"Thank you [chair name]. I'm [author] from [affiliation], and today I'll be talking about [short title]."

---

## Slide 2: Motivation [0:15 - 1:30]

"Let me start with the problem. [Describe the real-world motivation in accessible terms]. This matters because [impact statement].

The current state of the art approaches this with [brief existing approach]. But there's a fundamental limitation: [gap statement]."

→ *Transition*: "So what's our key insight?"

---

## Slide 3: Key Insight [1:30 - 2:30]

"Our key observation is that [core insight in one sentence].

This leads us to propose [method name], which [one-sentence description]."

→ *Transition*: "Let me walk you through how this works."

---

## Slide 4-N: [Continue for each slide...]

...

---

## Slide [N]: Thank You [TALK_MINUTES:00]

"To summarize: we've shown that [main result]. The key takeaway is [memorable final message].

The paper and code are available at the QR code on screen. I'm happy to take questions."

---

## Time Budget Summary

| Slide | Topic | Duration | Cumulative |
|:-----:|-------|:--------:|:----------:|
| 1 | Title | 0:15 | 0:15 |
| 2 | Motivation | 1:15 | 1:30 |
| 3 | Key Insight | 1:00 | 2:30 |
| ... | ... | ... | ... |
| N | Thank You | 0:15 | [TALK_MINUTES]:00 |

**Total**: [sum] min (target: [TALK_MINUTES] min)

---

## Anticipated Q&A

### Q1: How does this compare to [strongest baseline]?
**A**: "[Specific comparison with numbers]. Our advantage is particularly clear in [specific scenario], where we see [X%] improvement."

### Q2: What are the main limitations?
**A**: "[Honest answer]. We see this as [future work direction]."

### Q3: How computationally expensive is this?
**A**: "[Training/inference cost]. Compared to [baseline], our method requires [comparison]."

### Q4: Does this generalize to [related domain]?
**A**: "[Answer based on paper's discussion section]."

### Q5: What's the most surprising finding?
**A**: "[Interesting insight from the experiments]."

### Q6: How sensitive is the method to [hyperparameter/design choice]?
**A**: "[Reference ablation study if available]."

### Q7: What's the next step for this research?
**A**: "[Future work from conclusion]."

### Q8: [Domain-specific question]
**A**: "[Answer]."

Final Output Summary

📊 Slide generation complete:
- Talk type: [TALK_TYPE] ([TALK_MINUTES] min) for [VENUE]
- Files:
  slides/
  ├── main.tex              # Beamer LaTeX source
  ├── main.pdf              # Compiled slides (primary output)
  ├── presentation.pptx     # Editable PowerPoint
  ├── SLIDE_OUTLINE.md      # Slide-by-slide outline
  ├── SLIDES_REVIEW.md      # GPT-5.4 review feedback
  ├── speaker_notes.md      # Per-slide speaker notes
  ├── TALK_SCRIPT.md        # Full word-for-word talk script + Q&A
  ├── SLIDES_STATE.json     # State persistence
  ├── generate_pptx.py      # PPTX generation script
  └── figures/              # Symlinked from paper/figures/

Next steps:
1. Practice with TALK_SCRIPT.md (read aloud, time yourself)
2. Edit presentation.pptx for visual tweaks (animations, custom graphics)
3. Review Anticipated Q&A section before the talk
4. Do a dry run with a colleague

State: Write SLIDES_STATE.json with phase: 8, status: "completed".

Recommended Follow-up: /slides-polish

After this skill produces the initial Beamer + PPTX, the typical drift is typography proportion + per-slide layout, not content. Run /slides-polish as a focused post-generation polish phase: it does per-page Codex review against a reference visual (e.g., a prior academic talk), bumps PPTX fonts to projector-readable sizes, fixes text-frame overflow, and applies a fix-pattern catalog (italic style leaks, em-dash spacing, image aspect ratio, Chinese-font hints, anonymity placeholders).

/slides-polish is read-only on content (no claim / number / citation edits) and preserves speaker notes verbatim. Invocation:

/slides-polish slides/ — reference: <ref-pdf> [— style: generic | why-rf | <venue>]

Skip it for short decks (< 5 slides) or when a complete redesign is needed (re-run /paper-slides instead).

Key Rules

• Large file handling: If the Write tool fails due to file size, immediately retry using Bash (cat << 'EOF' > file) to write in chunks. Do NOT ask the user for permission — just do it silently.
• One message per slide. If a slide has two ideas, split it into two slides.
• Do NOT fabricate data. All numbers must come from paper/sections/*.tex.
• Bullet points only — never full sentences on slides. Sentence fragments are fine.
• Figure slides: figure ≥60% of slide area. The figure IS the content.
• Progressive disclosure: Use \pause or \onslide for complex method slides.
• De-AI polish: Remove watch words from all slide text and talk script.
• Do NOT hallucinate citations. Reference only papers cited in the paper.
• Opening hook matters: Never start with "In this paper, we..." — start with the problem or a provocative question.
• Font size minimums: Title ≥28pt, body ≥20pt, footnotes ≥14pt.
• Feishu notifications are optional. If ~/.claude/feishu.json exists, send notifications. If absent, skip.

Parameter Pass-Through

/paper-slides "paper/" — talk_type: oral, venue: ICML, minutes: 20, aspect: 4:3, notes: false

Parameter	Default	Description
venue	NeurIPS	Conference for color scheme
talk_type	spotlight	oral/spotlight/poster-talk/invited
minutes	15	Talk duration
aspect	16:9	Aspect ratio (16:9 / 4:3)
notes	true	Generate speaker notes
engine	pdflatex	LaTeX engine
auto proceed	false	Skip checkpoints