📜 Papyrus

Turn arXiv papers into bilingual, beautifully formatted, annotation-rich PDFs.

Papyrus (莎草纸) transforms academic papers into deep-read study documents. Given a paper URL, it produces a single PDF containing the original English text, paragraph-by-paragraph Chinese translation, expert commentary, all figures and formulas — typeset with professional Kami-inspired design.

---

✨ What You Get

Feature	Description
Original text preserved	Exact English wording, figures, tables, equations
Chinese translation	Paragraph-by-paragraph, accurate and idiomatic
Expert commentary	Technical insights, historical context, design rationale
Crisp formulas	LaTeX rendered at 150 DPI via codecogs, zero blur
Visible formula wireframes	Clean border + background for each equation block
Professional typesetting	Kami design system — 10pt MingLiU + Times New Roman
Appendix figures	All attention visualizations and supplementary charts

🚀 Quick Start

# CLI entry point
./SCRIPTS/run.sh --paper https://arxiv.org/abs/1706.03762 \
    --title "Attention Is All You Need — 逐段精读"

The pipeline:

1. Fetch — Downloads arXiv LaTeX source
2. Render — Converts LaTeX formulas to PNG via codecogs.com
3. Build — Constructs annotated bilingual HTML (LLM-assisted)
4. Export — WeasyPrint HTML → PDF

⚠️ Steps 2 (formula identification) and 3 (translation/commentary) require LLM assistance. Papyrus provides the scaffolding; the LLM fills in content following SOP.md.

🤖 Agent Platform Support

Papyrus v0.2.0 ships with adapters for major AI coding agents:

Platform	Config	How to Install
OpenClaw	SKILL.md (auto-loaded)	Already installed at ~/.openclaw/skills/papyrus/
Claude Code	platforms/claude-code/papyrus-skill.md	Copy to .claude/skills/
Codex	platforms/codex/papyrus-tool.yaml	Copy to .codex/tools/
Hermes	platforms/hermes/papyrus-tool.py	Copy to hermes/tools/
Open Code	platforms/open-code/papyrus-config.yaml	Copy to .open-code/

All platforms invoke the same unified CLI (SCRIPTS/papyrus) with four commands: fetch, figures, formulas, pdf.

📁 Structure

papyrus/
├── SKILL.md              # Skill definition and design decisions
├── SOP.md                # Standard Operating Procedure (mandatory)
├── README.md             # You are here
├── LICENSE               # MIT
├── SCRIPTS/
│   ├── run.sh            # Main entry point
│   ├── fetch_arxiv.sh    # Download arXiv LaTeX source
│   ├── render_formulas.sh # Render LaTeX → PNG
│   ├── build_html.py     # Build annotated HTML skeleton
│   └── build_pdf.sh      # HTML → PDF via WeasyPrint
├── TEMPLATES/
│   └── paper.html        # Kami-styled HTML template
└── PROMPTS/
    └── qc_checklist.md    # Three-round quality control

🔧 Dependencies

• Python 3.10+ — weasyprint, pypdf, Pillow, requests
• bash — standard Unix tools
• Option A (recommended): local LaTeX engine (pdflatex from texlive or basictex)
• Option B (backup): internet access for codecogs.com formula rendering

Install Python dependencies:

pip3 install weasyprint pypdf Pillow requests

For best results, install a lightweight LaTeX distribution:

# macOS
brew install --cask basictex

# Ubuntu/Debian
apt install texlive-latex-base texlive-latex-extra

🎨 Design Philosophy

• No scaling, no blur. Formula images use native 150 DPI resolution with max-width:100%; height:auto. Never upscale.
• Visible structure. Formula blocks have a wireframe box to visually separate math from text.
• Three-layer depth. Each paragraph: original English → Chinese translation → expert commentary.
• Defensive figure mapping. Always verify figure-to-file mapping from arXiv HTML source. Never guess.

📚 Example Output

See examples/ for a complete output: "Attention Is All You Need — Annotated Deep-Read" (38 pages, bilingual, 7 formulas, 5 figures).

🤝 Contributing

Found a rendering bug? Have a paper that Papyrus struggles with? Open an issue with:

1. The paper URL
2. A description of the problem
3. Screenshots of the issue

📄 License

MIT — see LICENSE.

---

Papyrus was born from the observation that reading academic papers shouldn't require fighting with layout engines. One session with "Attention Is All You Need" produced enough learnings to codify the entire pipeline.

---

🎨 Credits

The visual design system (warm parchment background, ink-blue accent, serif-led typographic scale) is adapted from the Kami design language, part of the OpenClaw skills ecosystem.

Kami's design tokens (colors, spacing, font stacks, wireframe patterns) are used under the same MIT license.