Paper2PR

A Claude Code multi-agent workflow that converts AI/ML papers into presentation-ready Beamer + Quarto (RevealJS) slide decks with automated quality control, adversarial review, and deployment.

Organization: WoRV / MaumAI

---

Papers

Paper	Topic	Beamer	Quarto
DreamZero	World Action Models as Zero-shot Policies (NVIDIA)	Slides/DreamZero.tex	Quarto/DreamZero.qmd
DreamDojo	A Generalist Robot World Model from Large-Scale Human Videos (NVIDIA)	Slides/DreamDojo.tex	Quarto/DreamDojo.qmd

Each paper gets a ~30-slide deck (~30 min) covering main ideas, technical details, and implementation observations. Target audience: basic deep learning knowledge.

---

Quick Start

Prerequisites

• Claude Code
• XeLaTeX (TeX Live / MacTeX)
• Quarto

Build

# Beamer (3-pass XeLaTeX)
cd Slides
TEXINPUTS=../Preambles:$TEXINPUTS xelatex -interaction=nonstopmode DreamZero.tex
BIBINPUTS=..:$BIBINPUTS bibtex DreamZero
TEXINPUTS=../Preambles:$TEXINPUTS xelatex -interaction=nonstopmode DreamZero.tex
TEXINPUTS=../Preambles:$TEXINPUTS xelatex -interaction=nonstopmode DreamZero.tex

# Quarto
cd Quarto && quarto render DreamZero.qmd

# Deploy to GitHub Pages
./scripts/sync_to_docs.sh DreamZero

With Claude Code

claude

Then use skills like /create-lecture, /translate-to-quarto, /slide-excellence, /deploy.

---

Project Structure

paper2pr/
├── CLAUDE.md                    # Project config for Claude Code
├── target-papers/               # Source papers (gitignored)
│   └── YYMM-papername/
│       ├── paper/               # LaTeX source, figures, bib
│       └── code/                # Official implementation
├── Bibliography_base.bib        # Centralized bibliography
├── Figures/                     # Per-paper subdirectories (PDF + SVG)
├── Preambles/header.tex         # Shared Beamer preamble
├── Slides/                      # Beamer .tex + compiled .pdf
├── Quarto/                      # RevealJS .qmd + theme
├── docs/                        # GitHub Pages (auto-generated)
├── scripts/                     # Build & deploy utilities
├── quality_reports/             # Plans, session logs, audit reports
├── guide/                       # Workflow guide (from upstream)
└── .claude/                     # Agents, skills, rules, hooks

---

Workflow

This project uses a multi-agent Claude Code workflow:

• 10 specialized agents — proofreader, slide-auditor, pedagogy-reviewer, domain-reviewer, tikz-reviewer, beamer-translator, quarto-critic, quarto-fixer, verifier, r-reviewer
• Adversarial QA — critic and fixer agents loop until approved (max 5 rounds)
• Quality gates — 80 (commit) / 90 (PR) / 95 (excellence)
• Plan-first — enter plan mode before non-trivial tasks
• Context survival — plans and session logs persist across compression

Key Skills

Command	What It Does
/create-lecture	Full lecture creation from paper
/translate-to-quarto	Beamer → Quarto translation
/slide-excellence	Combined multi-agent review
/compile-latex	3-pass XeLaTeX + bibtex
/deploy	Render Quarto + sync to docs/
/qa-quarto	Adversarial Quarto vs Beamer QA
/proofread	Grammar/typo/overflow review
/visual-audit	Slide layout audit

See CLAUDE.md for the full skill reference.

---

Attribution

This project is built on claude-code-my-workflow by Pedro H. C. Sant'Anna, originally developed for Econ 730 at Emory University. The multi-agent architecture, quality gates, orchestrator protocol, and infrastructure (agents, skills, rules, hooks) are from that foundation.

Thank you for making it open source.

---

License

MIT License. See LICENSE.