Research Hound: a scientific research workflow for Claude Code

Documentation: alveeerahman.github.io/research-hound · part of a three-skill suite with Skill Vision and Agent Oracle.

research-hound is an Agent Skill for Claude Code that turns Claude into a methodical research partner: a five-stage scientific workflow (critical thinking, literature review, brainstorming, writing, evaluation) backed by 21 runnable scripts that verify what a chat model would otherwise just assert. It hunts down the weak citation, the untraceable claim, and the statistical pitfall before a reviewer does.

Get it on the scent

# All your projects (personal skill):
git clone https://github.com/AlveeeRahman/research-hound.git ~/.claude/skills/research-hound

# Or one project only (shareable with your team through git):
git clone https://github.com/AlveeeRahman/research-hound.git .claude/skills/research-hound

Python 3.10+, standard library only, except requests in the one script that reaches the network: scripts/literature-review/verify_citations.py, which queries DOI resolvers to check a reference actually resolves. No script needs an API key, and nothing sends your prose, data or figures to a hosted model. Then ask Claude in your own words:

"Search the literature on X and verify every citation." "Audit the claims in my methods section against my sources." "Is this analysis p-hacked?" "Score this manuscript against the rubric and report inter-rater agreement."

The hound's route: five stages, twenty-one checks

A research question goes in one end. A manuscript that can survive review comes out the other. No stage is skippable, and every stage past the first carries its own runnable tooling:

flowchart LR
    q[/"❓ research question"/] --> ct
    subgraph THINK["Interrogate, before you search"]
        ct["1 · critical thinking<br/>evidence hierarchy · bias & fallacy screens"]
    end
    subgraph GATHER["Gather & shape"]
        direction LR
        lr["2 · literature review<br/>3 scripts · PubMed / arXiv / S2<br/>+ citation verification"] --> bs["3 · brainstorming<br/>3 scripts · scored ideation triage"]
    end
    subgraph SHIP["Write & judge"]
        direction LR
        wr["4 · writing<br/>8 scripts · IMRaD scaffold →<br/>claim audit → manuscript lint"] --> ev["5 · evaluation<br/>7 scripts · rubric scores +<br/>inter-rater agreement"]
    end
    ct --> lr
    bs --> wr
    ev --> out{{"📄 work that survives review"}}
    classDef think fill:#f59e0b,stroke:#b45309,color:#1f2937
    classDef gather fill:#3b82f6,stroke:#1d4ed8,color:#ffffff
    classDef ship fill:#22c55e,stroke:#15803d,color:#1f2937
    class ct think
    class lr,bs gather
    class wr,ev ship

How research-hound differs: the validation

vs. Claude's built-in research

Claude's native web search and Deep Research are retrieval and synthesis: they find sources and write summaries, in one pass, inside the chat. research-hound is a methodology harness on top of the model:

• Verification, not just citation. verify_citations.py checks that references actually resolve and audit_claims.py traces each manuscript claim to its evidence. Those are the two failure modes chat-only research cannot catch in itself.
• Staged discipline. Question interrogation before searching, searching before synthesis, a source ledger throughout, and rubric scoring with inter-rater agreement at the end. Built-in research has no stages to hold it accountable.
• Deterministic and reproducible. Every check is a plain CLI your CI can re-run. A conversation cannot be re-run. pytest-style verification can.
• Scientific-writing integrity rules. Formatting guidance keeps equations, symbols, units, subscripts, and superscripts intact through drafting and format conversion, and the skill deliberately ships no boilerplate LaTeX template. Templates that fabricate polish were removed in favor of integrity rules (see references/writing/professional_report_formatting.md).
• Research-integrity guardrails. Authorship and AI-confidentiality policy, responsible-AI brainstorming constraints, and reporting-guideline selection (CONSORT / PRISMA / STROBE) are first-class steps, not afterthoughts.

vs. the top research skills on GitHub

Surveyed August 2026 against the most-starred repos in the class (this niche's leaders sit at 1-2 stars, so the field is young):

Honest credit where due: chgagne's bibliography tooling is real and its stdlib-guard-test discipline is excellent, and jamoeight/deep-research-v2 is the strongest search orchestration in the class. Neither covers the lifecycle. What happens before the search (interrogating the question) and after it (writing, auditing, and scoring the result) is where research-hound lives. Where a lighter skill fits better, say a one-shot digest or a single-domain screen, use the lighter skill. This one is for work that has to survive review.

claude.ai / Desktop upload caveat

The frontmatter description is ~600 characters, tuned for Claude Code triggering. The claude.ai web uploader caps descriptions at 200 characters, so shorten it in your zip copy before uploading, and exclude .git:

zip -r research-hound.zip research-hound -x "research-hound/.git/*" "research-hound/.git"

What's new in v1.2.0

The description now fits claude.ai's uploader. It caps description at 200 characters; the Agent Skills spec allows 1024, so a 652-character description passed every local check and would still have been rejected on upload. It is 192 characters now, rewritten rather than truncated — a plain cut would have removed the whole "Use for…" clause, which is the half that decides when the skill triggers. The compatibility field was also unquoted while containing : , which made the frontmatter fail a strict YAML parse.

The image-generation stage is gone. It posted your diagram prompt to a hosted model and needed an API key, and its documentation had described one of its two scripts as an offline alternative when it was a thin wrapper around the online one. Rather than document that more loudly, the whole stage was removed: a research skill that uploads descriptions of unpublished work is the wrong default however clearly the upload is disclosed.

One script now reaches the network — verify_citations.py, checking that a DOI resolves — and no script needs an API key. Write diagrams as Mermaid inline; Claude Code and claude.ai both render it, which covers flowcharts, CONSORT and PRISMA diagrams, pathways and most mechanism figures.

Every reference file is now one link hop from SKILL.md. All 36 sat two hops out, behind a guide. Anthropic's authoring guidance names that as an anti-pattern, because past one hop an agent tends to preview a file with head -100 rather than read it, so the end is never seen. The guides still carry each stage's workflow. A compact index in SKILL.md links every reference directly, so either route works.

Tables of contents on all 30 reference files over 100 lines. A partial read now still shows the full scope of what is in the file.

Documented commands now run from where the docs say to run them. Several reference files invoked bare filenames like python search_databases.py for scripts that live under scripts/literature-review/. An agent following those literally fails. A stale block of upstream CI calling skills-ref, skill-scanner and scan_pr_skills.py, none of which ship here, is now labelled as the upstream project's pipeline instead of reading as instructions.

Licensing. The guides and references derive from K-Dense AI's MIT-licensed scientific-agent-skills, and MIT requires the original notice travel with them. LICENSE carried only one copyright line. It now retains both.

Description trimmed from 932 characters to roughly 600. That text loads into the system prompt of every session whether or not the skill ever triggers, which makes it the one cost paid unconditionally. It went from 233 tokens to 168.

CI now audits documentation claims against the code on every push, and sweeps every CLI for a working --help.

What's in the box

• SKILL.md: the five-stage routing instructions Claude loads (~3.6k tokens total context cost).
• guides/: one guide per stage.
• references/: 36 stage-organized reference documents (evidence hierarchies, database strategies, citation styles, IMRaD structure, evaluation frameworks…), each linked directly from SKILL.md.
• scripts/: 21 CLIs across four stages (critical thinking ships no scripts, by design).
• scripts/_shared/safe_io.py: the file-I/O module shared by the brainstorming, evaluation, and writing stages. It opens local files with O_NOFOLLOW and checks the resulting descriptor via os.fstat, so a symlink swapped in between a check and a read cannot be followed. Each stage keeps its own extra checks on top: duplicate-key rejection, non-finite-number rejection, structure and depth bounds, and, for evaluation, rejection of private-field keys.
• assets/: supporting fixtures.

License

MIT, copyright (c) 2026 MrPirate.

Composed from five skills in K-Dense AI's scientific-agent-skills (MIT): scientific-critical-thinking, literature-review, scientific-brainstorming, scientific-writing, and scholar-evaluation. The guide and reference bodies are derived from those originals and carry their upstream copyright, retained in LICENSE and detailed in NOTICE as MIT requires.

May your citations always resolve on the first fetch.