Read an arXiv paper (deep report)

Output locations (pkbllm convention)

• Tracked output (write here): $HUMAN_MATERIAL_PATH/research/<paper_slug>/
  • report.md (main deliverable)
  • figures/ (optional images/diagrams)
• Local downloads (gitignored, configurable): $HUMAN_MATERIAL_PATH/.references/
  • PDFs default: .references/pdfs/
  • Sources default: .references/arxiv/
  • Cloned repos default: .references/repos/

Config precedence (repo overrides user):

1. $HUMAN_MATERIAL_PATH/.agents/config.toml
2. ~/.agents/config.toml

Quick start

1. Download paper assets:

# Run from this skill directory (the folder containing this SKILL.md):
python scripts/download_arxiv.py https://arxiv.org/abs/2601.07372

2. Extract the TeX source:

python scripts/extract_arxiv_source.py 2601.07372

3. Write the report:

• Create $HUMAN_MATERIAL_PATH/research/<paper_slug>/
• Start from assets/report_template.md
• Fill it with a coherent story:
  • What problem is being solved and why now?
  • What is the key trick (the one idea that makes it work)?
  • What are the assumptions and failure modes?
  • What to implement first if we were to reproduce it?

Workflow (recommended)

1. Normalize the paper id from the URL (keep the version if present; it’s fine).
2. Download PDF + source into .references/ using the provided scripts.
3. Unpack sources into .references/arxiv/<arxiv_id>/ and locate the LaTeX entrypoint (main.tex, etc.).
4. Read in passes:
   • Pass 1: abstract + intro + figures (build the mental model)
   • Pass 2: method (write down the algorithm precisely)
   • Pass 3: experiments (what matters, what doesn’t)
   • Pass 4: limitations + related work (what breaks, what’s missing)
5. Produce a report under $HUMAN_MATERIAL_PATH/research/<paper_slug>/report.md:
   • Use the template (copy it and fill it)
   • Prefer concrete examples, small equations, and “why this design” explanations
   • Include at least one visual explanation if it helps (optionally generate via uv-scientific-schematics)