research-kit

A reusable Claude Code skill for deep research on any topic: fan-out web research with verification, then automatically find, download, and parse papers and explore public datasets — producing a structured, cited research folder you can build on.

It generalizes a pipeline first built for an ML project (menstrual-cycle prediction): the same downloaders that pulled 14 papers (PDF + arXiv LaTeX source) and catalogued/fetched public datasets, packaged so you can point it at a new question and get the same output.

What it produces

Run it on a topic and it writes a research folder:

<topic>/
  RESEARCH.md        # verified, cited synthesis (claims + sources + caveats)
  papers.csv         # catalogue: title, authors, year, venue, doi, url
  papers/            # downloaded PDFs (+ arXiv .tar.gz LaTeX source for math)
  DATASETS.md        # public datasets: scale, signals, access, licensing
  data/              # freely-downloadable datasets

Install

git clone https://github.com/winterchill-xyz/research-kit
cd research-kit && ./install.sh      # symlinks the skill into ~/.claude/skills/
cp .env.example .env                 # add credentials (or point at an existing .env)

Then in Claude Code:

/research-kit <your research question>

Credentials

The downloaders use two services (keys go in .env, or pass --env <path>):

• BrightData Web Unlocker — fetches paywalled/blocked PDFs and dataset landing pages. Needs the runner's egress IP whitelisted in the zone, and to run outside a network sandbox (see references/sources.md).
• OpenRouter — Perplexity models for discovering a direct PDF link when the open-access routes miss.

Open-access resolution (arXiv, Europe PMC, OpenAlex, Unpaywall) needs no keys — only a contact email for Unpaywall.

Layout

Path	What
SKILL.md	the skill definition + the research workflow Claude follows
scripts/	topic-agnostic downloaders (fetch_pdfs, fetch_arxiv_html, fetch_datasets)
templates/	starters for RESEARCH.md, papers.csv, DATASETS.md
references/sources.md	OA resolution routes + BrightData/OpenRouter usage notes

MIT licensed. Downloaded papers/datasets are for private research; respect each source's license before redistributing.