arxiv-reading

name: arxiv-reading description: Use this skill when the user provides an arXiv ID (e.g., arXiv:2508.12856) and asks to read, analyze, or extract equations from the paper. Activates for literature review, equation extraction, paper understanding, and cross-referencing tasks in academic research contexts.

arXiv Literature Reading Protocol

Overview

When an arXiv ID is provided, use the scripts in scripts/ to download, extract, and analyze the paper. Write a memory file <user-working-dir>/arxiv-reading/<arxiv-id>.memory.md for further checking.

IMPORTANT — Cache directory is the user's working directory: All cache files live under <user-working-dir>/arxiv-reading/ — NOT the skill's directory. The user's working directory is the "Primary working directory" shown in the system context (session working directory), not the skill's installation directory.

Always construct absolute paths manually using the known working directory from the conversation context.The shell's $PWD and Python's os.getcwd() are unreliable because the skill may run from its own directory.

Script-based Workflow

0. Check for cached data

Before downloading, check if the memory file <user-working-dir>/arxiv-reading/<arxiv-id>.memory.md already exists. If it does, directly read it and find information. If you need more information, check individual sections at <arxiv-id>.sections_content.json in the same directory.

Example:

memory_file="<user-working-dir>/arxiv-reading/<arxiv-id>.memory.html"
if [ -f "$cache_file" ]; then
  echo "Cache hit: $cache_file"
fi

1. Download ar5iv HTML (only if not cached)

If no cached file exists, use the skill's fetch.py to download the ar5iv HTML version. Always pass --out-dir with the absolute path to <user-working-dir>/arxiv-reading/.

• Primary Source: https://ar5iv.labs.arxiv.org/html/<arxiv-id>
• Uses requests; falls back to curl.
• Output is always written to <user-working-dir>/arxiv-reading/<arxiv-id>.html.
• Call with: python3 <skill-dir>/scripts/fetch.py --out-dir <user-working-dir>/arxiv-reading <arxiv-id>

Where <skill-dir> is ~/.claude/skills/arxiv-reading/ and <user-working-dir> is the session's primary working directory.

2. Equation Extraction

Use extract_eqs.py to extract all LaTeX equations with their numbers and surrounding context.

Call with the absolute path to the cached HTML:

python3 <skill-dir>/scripts/extract_eqs.py <user-working-dir>/arxiv-reading/<arxiv-id>.html

Auto-caching: On first run, saves JSON to <arxiv-id>.eqs.json. Subsequent runs read from cache directly.

3. Section Extraction (for targeted reading)

Use extract_sections.py to list all sections or to extract a specific section's content as clean plain text.

Call with the absolute path to the cached HTML:

python3 <skill-dir>/scripts/extract_sections.py <user-working-dir>/arxiv-reading/<arxiv-id>.html
python3 <skill-dir>/scripts/extract_sections.py <user-working-dir>/arxiv-reading/<arxiv-id>.html "Section Name"

Auto-caching: Section listing is cached as <arxiv-id>.sections.json. On first content extraction, ALL sections' content is cached as <arxiv-id>.sections_content.json. Future reads (even for different sections) use cache.

4. LaTeX to Plain Text (for equation comprehension)

Use latex_simplify.py to convert a LaTeX math string into readable plain text (Greek unicode, simplified fractions, etc.).

Auto-caching: Results are cached in ~/.cache/arxiv-reading/latex_simplify_cache.json keyed by MD5 hash. Repeated calls with the same LaTeX string return instantly.

5. Memory file

Generate memory file <user-working-dir>/arxiv-reading/<arxiv-id>.memory.md for further checking.

File size: as a memory file for you to read comfortably.

• Summary per section & subsection: Write a summary of each section and subsection (if exists). So that you can find which section to further read in need.
• Equations: Note important equations (use latex_simplify.py to convert latex symbols to utf8 for readability) with their equation number. The user may refer to the equation number and you need to find the equation. Write a brief description of each equation. For short papers (the largest equation number is less than 15), note down all equations.
• References: Note down main references for further studies. Important: note down the title and arxiv number arXiv:xxxx.xxxxx of each reference (if available).
• Related to research mission: If you know the purpose of the current research, note down more details related to the current research.

Verification

1. Run extract_eqs.py to get all equations with their numbers.
2. Cache file <arxiv-id>.eqs.json is created alongside the HTML.
3. Cross-reference extracted formulas with the paper's internal equation numbering.
4. If numbering seems off, delete the .eqs.json cache and re-run to force re-extraction.
5. The memory file .memory.md is generated.

Available scripts

Scripts are located at the skill's directory (~/.claude/skills/arxiv-reading/scripts/). Always reference them by absolute path.

Cache files

All cache files live under <user-working-dir>/arxiv-reading/. The latex_simplify_cache.json lives in ~/.cache/arxiv-reading/ (shared across all projects).

Cache invalidation

To force re-extraction from HTML, delete the corresponding cache file (e.g., rm <arxiv-id>.eqs.json). The scripts will re-extract and re-cache on next run.

Notes

• All cache files are read/written under <user-working-dir>/arxiv-reading/ — the arxiv-reading/ directory in the user's session working directory, not the skill's installation directory.
• The shell environment $PWD and Python os.getcwd() may point to the skill's directory. Do not rely on them. Always substitute <user-working-dir> with the absolute path from the conversation context.
• If requests is not installed, use curl as a fallback:

  mkdir -p "<user-working-dir>/arxiv-reading"
  curl -sL "https://ar5iv.labs.arxiv.org/html/<id>" -o "<user-working-dir>/arxiv-reading/<id>.html"
  

• For PDF-only papers (no ar5iv), inform the user and suggest an alternative download method.
• ar5iv HTML files contain massive MathML markup. Running grep on them produces enormous output that gets persisted to disk, triggering permission prompts.