MedSci Skills is a set of Agent Skills for clinical research: literature and references, study design, statistics, figures, manuscript drafting, reporting-guideline checks and journal submission, plus a lane for building and validating medical-imaging AI models.
It is for physicians and biomedical or medical-engineering researchers who work in Claude Code, Codex, Cursor or GitHub Copilot.
The skills draft and check, bundled scripts recompute what can be recomputed, and every output still needs review by a qualified researcher; it is not a diagnostic tool or an autonomous author.
Installation
Without a terminal (Windows or macOS): download the classroom installer, unzip it and double-click the installer inside. It also turns on update reminders and puts an Update MedSci Skills icon on your Desktop. If you have not installed Claude Code, Python or Node yet, the setup guides for Mac and Windows go step by step.
In a terminal, with Node 18+ and Python 3.9+ installed:
npx medsci-skills install
This copies every skill into ~/.claude/skills/ (read by Claude Code, Cursor and Copilot in VS Code) and ~/.agents/skills/ (read by Codex, Cursor and GitHub Copilot).
Restart your agent, type /orchestrate, and describe what you want to do; it routes the request to the right skill.
To be told when a new version ships, add --enable-update-notify: a one-line notice at Claude Code session start, off by default, no telemetry.
Other ways to install, each described in docs/install.md:
Updating, what individual skills need (pandoc, R, PyTorch), where the files go and the optional routing block for plain-language requests are on the same page.
Three workflows to start with
Call a skill by name, or describe the task to /orchestrate.
Check a manuscript against its reporting guideline — /check-reporting
- You give: your manuscript and, if you know it, the guideline — it covers 49 reporting guidelines and risk-of-bias tools, from STROBE, STARD and CONSORT to PRISMA 2020 and TRIPOD+AI.
- You get: an item-by-item audit (PRESENT, PARTIAL, MISSING or N/A, with where each item was found), a list of fixes, and a JSON summary in
qc/; it is a working audit, not the checklist you upload to the journal.
Analyse a dataset — /analyze-stats
- You give: a de-identified data file (CSV, Excel or TSV) and your research question.
- You get: an analysis plan to approve, then Python (or R) code that it runs, tables as CSV and Markdown, figures as PDF and 300-dpi PNG, and a manifest that
/make-figures and /write-paper read.
Verify references — /verify-refs
- You give: a manuscript or bibliography (
.md, .docx, .bib, .txt or .tsv).
- You get:
qc/reference_audit.json, marking each reference OK, MISMATCH, UNVERIFIED or FABRICATED after lookup in PubMed, CrossRef and OpenAlex; it reports and never edits your references.
Longer chains (pre-submission audit, data to manuscript, systematic review) and step-by-step command lists for common tasks are in docs/workflows.md.
Demos
Five worked examples on public datasets, with their code, outputs and QC reports committed to the repository:
| Demo |
Data |
What it shows |
| Diagnostic accuracy |
Wisconsin breast cancer (sklearn) |
Analysis to manuscript draft, STARD 2015 audit |
| Meta-analysis |
BCG vaccine trials (metafor::dat.bcg) |
Pooled analysis to manuscript draft, PRISMA 2020 audit |
| Survey epidemiology |
NHANES 2017–18 |
Weighted survey analysis to manuscript draft, STROBE audit |
| Model engineering |
PneumoniaMNIST |
CNN scaffold, leakage gates, training, evaluation, Grad-CAM |
| External validation |
MSD spleen, then AMOS CT and MRI |
3-D segmentation tested on an external cohort, including where it failed |
What each demo produced, and how to rerun it: docs/demos.md.
Skills
All 54 skills, by research stage. Each name links to its reference page (what it does, when to use it, its known limits). npx medsci-skills list prints the same groups, and with the plugin install each row is one plugin.
Not sure which one fits? Start with /orchestrate.
Renamed in v6 (the old names still work until v7): imaging-data ← preprocess-imaging, profile-imaging; model-assessment ← explainability, model-evaluation, model-validation, uncertainty-imaging; model-selection ← architecture-zoo, model-sourcing.
Patient data and safety
Do not give an agent identifiable patient data. /deidentify runs locally with no network or AI calls: it detects protected health information with regex and heuristics (locale packs for twelve countries) and pseudonymises it after an interactive review, and /analyze-stats asks whether a raw data file contains patient identifiers before using it. These are research productivity tools, not clinical decision support: they are not clinically validated and do not replace expert review, so a qualified researcher must check every output before it is used in a publication or clinical context. The 93 deterministic detectors recompute or cross-check specific things (reference metadata, arithmetic, checklist items, data leakage); a clean run means those checks found nothing, not that the manuscript is correct. MEDSCI_AUDIT.md lists each detector and which have been formally evaluated; reference lookups use the public, keyless APIs in docs/connectors.md.
Citation
If MedSci Skills helped produce your manuscript, protocol, or analysis, please cite it —
software citation is how a tool like this earns academic recognition, and it takes one line.
In your manuscript (Methods or Acknowledgements — cite the version you actually used):
Reporting-guideline compliance, reference verification, and pre-submission integrity checks
were assisted by MedSci Skills (version X.Y.Z; https://github.com/Aperivue/medsci-skills;
archived at Zenodo, https://doi.org/10.5281/zenodo.20155321).
BibTeX (the software, and the preprint describing its design):
@software{nam_medsci_skills,
author = {Nam, Yoojin},
title = {{MedSci Skills: Claude Code Skills for the Medical Research Lifecycle}},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.20155321},
url = {https://github.com/Aperivue/medsci-skills}
}
@article{nam2026agentic,
author = {Nam, Yoojin and Jeong, Jinhoon and Kim, Namkug},
title = {{Deterministic Integrity Gates for LLM-Assisted Clinical Manuscript
Preparation: An Auditable Biomedical Informatics Architecture}},
year = {2026},
journal = {arXiv preprint arXiv:2606.09500},
url = {https://arxiv.org/abs/2606.09500}
}
The Zenodo concept DOI 10.5281/zenodo.20155321
always resolves to the latest release; CITATION.cff carries the machine-readable
metadata (GitHub’s “Cite this repository” button reads it).
Used it in published or in-review work? Tell us via the
“Used in research” issue template
— with your permission it is added to docs/citations.md.
License
MIT License. See LICENSE for details.
Some bundled material is not ours and is not MIT: the official guideline templates, the CSL citation styles, and a few checklist summaries carry their own terms — including CC BY-NC, which restricts commercial use. Those are indexed in THIRD-PARTY-NOTICES.md, which ships with every copy and is checked against the tree on every build.
Reporting-guideline checklists keep their sources’ terms, and several have no open licence, so those are included only as summaries in our own words; check-reporting’s LICENSES.md records each one.
Optional dependency: pdf_to_md.py uses pymupdf4llm (AGPL-3.0). Not bundled – installed separately by the user via pip install pymupdf4llm.
Acknowledgements
make-figures Critic Loop is inspired by PaperBanana (Zhu et al., Automating Academic Illustration for AI Scientists, arXiv:2601.23265, 2025) and by prior self-refinement research — Self-Refine (Madaan et al., 2023), Reflexion (Shinn et al., 2023), and Constitutional AI (Anthropic, 2022). The implementation in this repository is a clean-room reconstruction specialized for medical publication figures; no code, prompts, or configurations are derived from PaperBanana’s repository.- Reporting-guideline checklists bundled with
check-reporting credit their original authors in each file’s header.
- Wong colorblind-safe palette: Wong B. Points of view: Color blindness. Nature Methods 8:441 (2011).
More
- Release notes — what changed in each version.
- Upgrading from v5 — renamed skills, and how to update for each install channel.
- Workflows and skill boundaries — skill chains, which skill to use when two look alike, and checks that span skills.
- Skill reference and FAQ.
- Host compatibility — verified install paths for Claude Code, Codex, Cursor and GitHub Copilot.
- How it differs from other skill collections and the scope boundary.
- Contributing — most contributions are one file;
/contribute sends a change from your installed copy after scanning it for patient data. Good first issues.
- Adoption (stars, forks, downloads) and citations.
- Governance: ROADMAP, MAINTAINERS, SECURITY, CONTRIBUTORS.
Built by Aperivue.
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?v=4)
评论