Turn a reading list into source-grounded notes, check a draft against them, and build a paper.

Catalog · Install guide · Workflow · Changelog

Early-stage plugin 0.1.1: 7 skills, 4 agent roles, 8 command-line tools and project, Python and R templates. Install the whole Claude Code plugin or take the parts you need. The skills and roles are Markdown instructions; the helpers are seven Python programs and one POSIX shell script. Reading a skill does not execute a tool.

Install

/plugin marketplace add nik110403/research-kit
/plugin install research-kit@research-kit

Then ask Claude to start a research project, search the literature, or verify notes. Command-line dependencies still need to be available; see INSTALL.md for tool paths, selective installation and existing-settings precautions.

What you get

Kind	Name	What it does
Skill	new-research-project	Scaffold a numbered project tree from a question
Skill	literature-search	Search providers, save a reading list with a receipt
Skill	pdf-to-markdown	Convert PDFs to page-marked Markdown
Skill	literature-notes	Write a structured note with page-pinned quotes
Skill	review-against-notes	Review a draft against the notes
Skill	paper-build	Check the bibliography and build PDF or HTML
Skill	analysis-pipeline	Set up and run a repeatable analysis
Agent	source-reader	Read one source into a note
Agent	claim-checker	Judge whether notes support cited claims
Agent	skeptic	Challenge the strongest claims of a draft
Agent	pipeline-runner	Run a trusted analysis pipeline and report
Tool	verify-quotes	Check quotations against converted source text
Tool	lit-search	Literature search with provider receipts
Tool	pdf2md	PDF to Markdown with page markers
Tool	check-claims	Find missing or thin notes behind cited claims
Tool	check-bib	Check bibliography structure and citation coverage
Tool	build-paper	Build the paper with pandoc (POSIX shell)
Tool	new-project	Scaffold a research project
Tool	new-pipeline	Scaffold an R or Python pipeline

The complete catalog lists every component's inputs, outputs and dependencies; the adapter guide covers manual use with other assistants.

30-second example

The transcript above runs the offline quote check on the bundled example. The sources are bundled, so it needs neither PDF software nor an API key.

To reproduce it, clone the repository and, with Python 3.9+, run from its root:

cd examples/basic
python3 ../../plugins/research-kit/bin/verify-quotes 01_notes/lit/*.md
python3 ../../plugins/research-kit/bin/verify-quotes 01_notes/negative/*.md

The first command should print 14 of 14 PASS and exit 0. The second should report the fabricated quotation as FAIL and exit 1; that failure is expected. The source conversions are bundled, so this check needs neither PDF software nor an API key. See the example walkthrough and source attribution. Quote matching checks converted text and page markers, not whether a claim follows from the quotation.

How it works

Enter at any stage for which you have the inputs; the tools handle mechanical checks and the assistant roles return text for the main session to save. Details and evidence boundaries are in docs/workflow.md.

flowchart LR
  A[Scaffold] --> B[Search]
  B --> C[Acquire PDFs]
  C --> D[Convert]
  D --> E[Read and verify quotes]
  E --> F[Draft]
  F --> G[Review against notes]
  G --> H[Bibliography check and build]
  A -.-> I[Analysis side track]
  I -.-> F

Loading

Configuration

• Python helpers require Python 3.9+ and use the standard library, except the optional Docling conversion backend. Search needs internet access. Semantic Scholar can reject keyless requests; an optional SEMANTIC_SCHOLAR_API_KEY helps, and receipts report provider failures and partial coverage.
• PDF conversion needs Poppler's pdftotext; scans may need ocrmypdf, encrypted files qpdf. Docling currently emits one block marked page 1, not original PDF pagination. Recheck page pins against the PDF.
• Paper builds need POSIX sh and pandoc; XeLaTeX enables PDF output, otherwise the tool builds HTML. On Windows, use Git Bash or WSL for build-paper. UTF-8 Python tests do not establish whole-workflow Windows support.
• R pipelines need R; Python pipelines can use uv. Running a pipeline executes project code with your permissions. Model steps also depend on your assistant's access and may incur its usual costs. No native non-Claude integration or cross-platform performance guarantee is claimed.

Contributing

CONTRIBUTING.md covers checks, privacy and licensing.

License

Code, skills, agents and templates are MIT. Documentation under docs/ is CC BY 4.0; bundled papers and conversions retain their authors' attribution. See the licensing map.