Aperivue/medsci-skills

MedSci Skills – What it is

MedSci Skills is an open‑source collection of Agent Skills that let a coding‑assistant (Claude Code, GitHub Copilot, Cursor, Codex, etc.) drive the whole lifecycle of a clinical research project – from finding the literature to producing a manuscript and, when relevant, building a reproducible medical‑AI model.

The repository ships 59 ready‑to‑use “skills” (small, self‑contained scripts) that can be installed into an AI‑coding agent. Each skill performs a specific, well‑defined step – e.g. search-lit, clean-data, analyze-stats, check-reporting, verify-refs, make‑figures, write-paper, or scaffold‑model. The skills are grouped into nine plugin categories (literature, data, modeling, analysis, writing, review, submission, project, presentation) so users can install just the parts they need.

Why it matters

• Compliance‑first – every output passes deterministic “integrity gates” that check against 49 reporting guidelines (STARD, PRISMA, STROBE, CLAIM, TRIPOD‑AI, etc.) and citation‑integrity tools. This makes the generated manuscript and model‑code suitable for submission‑grade biomedical research.
• Model‑engineering lane – starting with version 5 the toolkit can scaffold a full PyTorch training repository (using MONAI, nnU‑Net, timm, torchvision), enforce leakage‑safe data splits, generate model cards, and run a smoke‑test forward pass. It is not a diagnostic AI; it only helps researchers build and audit reproducible AI models.
• Agent‑centric – the skills are designed to be called directly from an AI coding assistant (/orchestrate, /analyze-stats, /medsci‑analysis:analyze‑stats, etc.). The same code works whether you install via npx medsci‑skills, GitHub CLI (gh skill), or the Claude Code plugin marketplace.
• End‑to‑end demos – the repo includes five complete pipelines (diagnostic accuracy, meta‑analysis, epidemiology, CNN segmentation, 3‑D external validation) that produce a manuscript, figures, compliance checklists, and, for the AI lanes, training logs and external‑validation reports.

Core components

Component	What it does	Example skill
Literature & discovery	Searches PubMed/Europe PMC, pulls full‑text, syncs with Zotero, audits references.	medsci-literature:search-lit
Data preparation	Designs study protocol, calculates sample size, de‑identifies PHI, cleans CSV/Parquet, builds codebooks.	clean-data
Statistical analysis	Runs R/Python stats, creates forest/ROC plots, produces reproducible notebooks.	analyze-stats
Manuscript drafting	Generates IMRaD sections, removes AI‑generated patterns, formats citations.	write-paper
Compliance & review	Checks against EQUATOR guidelines, runs self‑review scoring, flags missing items.	check-reporting, self-review
Model engineering	Picks a paper‑grounded architecture, scaffolds a PyTorch repo, runs leakage checks, produces Model Card/Datasheet.	scaffold-model, profile-imaging
Submission packaging	Bundles manuscript, figures, checklists, fills ICMJE/IRB forms, suggests journals.	sync-submission, find-journal
Presentation	Renders PPTX/PDF, publishes skill bundles.	present-paper

Typical user workflows

1. Manuscript pre‑submission audit – run /self-review → /check-reporting → /verify-refs → /sync-submission on a near‑final draft to get a reviewer‑style comment list and a compliance checklist.
2. Data‑to‑paper – start from a cleaned dataset, invoke /clean-data → /analyze-stats → /make-figures → /write-paper → /check-reporting → /find-journal to obtain analysis code, publication‑ready figures, a draft manuscript, and a shortlist of suitable journals.
3. Systematic review / meta‑analysis – chain /search-lit → /meta-analysis → /make-figures → /check-reporting to produce a PRISMA protocol, extraction tables, pooled estimates, and a draft manuscript.
4. Medical‑AI model study – use the model‑engineering lane: /choose-arch → /scaffold-model → /run‑training → /profile-imaging → /evaluate → /check-reporting to get a reproducible training repo, validation metrics, and a CLAIM‑compliant audit.

Installation options (quick)

• One‑liner (no git): npx medsci-skills install [--enable-update-notify]
• GitHub CLI (skill preview & selective install): gh skill install Aperivue/medsci-skills check-reporting
• Claude Code plugin marketplace: /plugin marketplace add Aperivue/medsci-skills
• Classroom ZIP installer for users who prefer a double‑click GUI.

Who built it?

Created and maintained by Yoojin Nam, MD, a physician‑researcher in radiology at the University of Ulsan (Seoul). The project is MIT‑licensed, published on npm (medsci-skills), and has a DOI (Zenodo) and an arXiv pre‑print (arXiv:2606.09500).

---

Bottom line: MedSci Skills is a genuine, open‑source software suite that equips AI coding agents to automate the research side of clinical AI – from literature review through compliant manuscript generation and reproducible model engineering. It is squarely in the AI/ML frontier for biomedical research.