Readable Paper

A portable skill for turning a mathematical paper into a source-linked argument map and notation inventory.

Use it with your agent

Unzip the package into your agent’s skills directory. For Codex, place the readable-paper folder in ~/.codex/skills/, then start a new session.

Use $readable-paper on this PDF. Map its results, audit notation, and explain the main proof with source links. Preserve unresolved references.

Run the extractor locally

Python 3.10+ is required. JSON input needs only Python’s standard library. PDF extraction additionally requires PyMuPDF; its licensing terms apply separately.

python -m pip install pymupdf
python readable-paper/scripts/build_map.py paper.pdf \
  --output paper-map.json --document-output document.json

This command makes no network or model calls. It creates a candidate map. An agent using the skill can then inspect the source and add explanations and corrections in a separate enrichment file.

Bring the result back

On Paper maps, open your map JSON and matching PDF. They stay in the browser; the SHA-256 check prevents pairing a map with the wrong manuscript. You can also keep the JSON in your own research folder.

What is included

• Reusable extraction and evidence-validation scripts.
• A schema guide for source-linked nodes, relations, notation, and agent explanations.
• Instructions to distinguish textual references, inferred dependencies, missing evidence, and formal verification.
• A permissive license for the skill’s original code and instructions.

The standalone skill does not include live AI chat, automatic retrieval of external papers, OCR, or Lean proof checking. The website’s optional assistant requires separate API configuration. Equations extracted from PDFs can lose symbols; the original page is the reference.

Explore sample maps →