🐕 Research Hound
A scientific research workflow for Claude Code. Five stages — critical thinking, literature review, brainstorming, writing, evaluation — backed by 21 runnable scripts that verify what a chat model would otherwise just assert.
GitHub repository Latest release
Install as a Claude Code skill
git clone https://github.com/AlveeeRahman/research-hound.git ~/.claude/skills/research-houndPython 3.10+, standard library only except the one script listed under what touches the network. Then ask Claude: "search the literature on X and verify every citation", "audit the claims in my methods section", "is this analysis p-hacked?", "score this manuscript against the rubric".
Five stages, twenty-one checks
| # | Stage | Runnable tooling |
|---|---|---|
| 1 | Critical thinking | Evidence hierarchy, logical fallacies, statistical pitfalls, bias catalogs |
| 2 | Literature review | Result aggregation, deduplication and ranking across PubMed / arXiv / Semantic Scholar, citation verification against DOI resolvers, PDF generation |
| 3 | Brainstorming | Scored ideation matrix, session scaffolds |
| 4 | Writing | IMRaD scaffold, manuscript lint, claim-evidence audit, reference checks, CONSORT / PRISMA / STROBE selection |
| 5 | Evaluation | Rubric validation and scoring, inter-rater agreement, weight sensitivity, traceability |
Claude runs the searches with its own web tools. search_databases.py is the local half — it deduplicates by DOI, ranks, filters by year, and emits JSON, Markdown or BibTeX, without opening a connection of its own. The script that does reach out is verify_citations.py, which checks each DOI actually resolves.
What touches the network
One of the 21 scripts opens a connection. Everything else runs offline on the standard library, and no script needs an API key.
| Script | Reaches | Why |
|---|---|---|
literature-review/verify_citations.py | doi.org, api.crossref.org | Confirms a DOI resolves and the metadata matches the citation. |
Nothing here sends your prose, data or figures to a hosted model. An earlier version bundled an image generator that posted your diagram prompt to a third-party API; it was removed rather than documented more loudly. Write diagrams as Mermaid inline — it renders in Claude Code and claude.ai and never leaves the machine.
Why not just use Claude's built-in research?
Built-in web search and Deep Research are retrieval and synthesis — one pass, inside the chat. Research Hound is a methodology harness: citations get verified rather than merely cited, manuscript claims are traced to evidence, and every check is a plain CLI your CI can re-run. It covers what happens before the search (interrogating the question) and after it (writing, auditing, and scoring the result).
Project resources
Part of a three-skill suite: Skill Vision · Agent Oracle · Research Hound (this page)