ArcDeck — Claude Code Plugin

Convert academic PDF papers into polished, narrative-driven PowerPoint presentations using 13 specialized AI agents orchestrated by Claude Code.

Based on Rhetorical Structure Theory (RST) discourse parsing with a multi-agent critique-revise-judge refinement loop and CARP+ aesthetic design principles. Every slide is visually rich — text-only slides get programmatic diagrams and charts rendered via PptxGenJS (Node.js).

Paper: Narrative-Driven Paper-to-Slide Generation via ArcDeck

Official Website: https://arcdeck.org

Official GitHub Repository: https://github.com/RehgLab/ArcDeck

Installation

From Plugin Directory

claude --plugin-dir ./

From Marketplace

/plugins install arcdeck

Usage

Once installed, invoke the skill:

/arcdeck:arcdeck paper.pdf

Full Options

/arcdeck:arcdeck paper.pdf --audience researchers --duration 20 --template ocean --speaker-notes

Arguments

Argument	Description	Default
<pdf_path>	Path to the input PDF paper	(required)
--audience <type>	Target audience: researchers, students, industry, general	researchers
--duration <minutes>	Target talk length in minutes	20
--template <theme>	Presentation theme: ocean, minimal, dark, warm	ocean
--speaker-notes	Generate speaker notes for each slide	off

Presentation Themes

Theme	Style	Colors
ocean	Ocean Gradient — navy/teal/cyan	#065A82 #1C7293 #0891B2
minimal	Clean Minimal — white/coral accent	#E63946 #457B9D #2A9D8F
dark	Dark Mode — charcoal/violet accent	#7F5AF0 #2CB67D #FF6B6B
warm	Warm Academic — cream/terracotta	#C45B3E #5B8A72 #D4A84B

Prerequisites

• Python 3.10+ with docling installed (PDF extraction)
• Node.js 18+ with pptxgenjs (PPTX rendering)
• LLM API key: OPENAI_API_KEY or ANTHROPIC_API_KEY set in environment

Install Python dependencies:

pip install docling Pillow

Node.js dependency is installed automatically in the workspace during execution.

How It Works

ArcDeck runs a 13-agent pipeline across 9 phases:

Phase	Agents	Purpose
0	—	Setup & user input
1	A1-A2	PDF preprocessing & asset extraction (Docling)
2	A3	Commitment building (global narrative contract)
3	A4	RST discourse parsing (per-section)
4	A5	Slide planning (RST-guided grouping)
5	A6-A8	Narrative refinement loop (critique-revise-judge, up to 3 rounds)
6	A9-A10	Visual asset filtering & slide composition
7	A11-A12	Aesthetic critique-refine (CARP+ design principles)
8	A13	PPTX assembly via PptxGenJS

Agents A10-A13 are Claude Code native — Claude Code acts as the LLM directly with no external API calls.

Plugin Structure

arcdeck-plugin/
├── .claude-plugin/
│   └── plugin.json              # Plugin manifest
├── skills/
│   └── arcdeck/
│       ├── SKILL.md             # Main skill definition
│       ├── agents/              # 13 agent specifications (A1-A13)
│       ├── prompts/             # Prompt templates (outline + pipeline)
│       ├── reference/           # Reference docs (RST, templates, troubleshooting)
│       └── scripts/             # Pipeline scripts (Python + Node.js)
└── README.md

Output

Generated files are saved under workspace/ in the current directory:

File	Description
workspace/output.pptx	Final PowerPoint presentation
workspace/speaker_notes.json	Speaker notes (if requested)
workspace/slide_plan_refined.json	Final slide plan (JSON)
workspace/figures/	Extracted figures and tables

Citation

@article{ozden2026arcdeck,
  title     = {Narrative-Driven Paper-to-Slide Generation via ArcDeck},
  author    = {Ozden, Tarik Can and VS, Sachidanand and Horoz, Furkan
               and Kara, Ozgur and Kim, Junho and Rehg, James M.},
  journal   = {arXiv preprint arXiv:2604.11969},
  year      = {2026}
}

License

MIT

Contact

For feedback, improvements, or support: tozden2@illinois.edu