Research-Grade Skills
CI-validated Agent Skills for science. 69 skills covering scientific writing, machine learning, data analysis, quantum computing, and research databases — every one structurally verified on every commit.
Works with Claude Code, Cursor, Codex, and any client that supports the open Agent Skills standard.
Browse the interactive catalog
Why This Exists
Researchers cannot afford silent errors. One wrong statistical method, one invented API, one outdated citation — and hours of work (or a paper) are compromised. Skill libraries optimized for size cannot make that guarantee; a library a researcher actually relies on has to earn it.
This repository is an independently maintained, heavily curated descendant of K-Dense's scientific skills. Compared to where it started:
- All vendor promotion removed. Upstream, every skill ended with a section steering the agent to advertise a commercial platform, plus one skill whose only job was self-promotion ("ALWAYS run this skill with every session"). All of it is gone, and CI now rejects any skill that tries to reintroduce vendor steering.
- A real quality gate. Every push and pull request validates all 69 skills: frontmatter schema, skill-name/directory consistency, registry-to-disk consistency, relative link resolution, and Python script syntax. A weekly job audits all external links for rot. None of this existed before.
- Bugs actually fixed. The validation gate already caught and fixed broken reference links, mismatched skill names, and an undefined-variable bug in a bundled script — the kind of defects that silently degrade agent behavior.
The goal is simple: any skill you load from here should be safe to use in real research output.
Install
Claude Code (plugin)
/plugin marketplace add Yuuqq/research-grade-skills
/plugin install scientific-skills@research-grade-skills
Cursor, Codex, and other Agent Skills clients
Clone the repository and copy the skills you need into your client's skills directory:
git clone https://github.com/Yuuqq/research-grade-skills.git
# Claude Code: ~/.claude/skills/
# Cursor: ~/.cursor/skills/ (or .cursor/skills/ in a project)
cp -r research-grade-skills/scientific-skills/literature-review ~/.cursor/skills/
Each skill is a self-contained folder — SKILL.md plus optional references/, scripts/, and assets/.
What's Inside
| Category | Skills |
|---|
| Scientific Communication (16) | literature review, peer review, scientific writing, hypothesis generation, citation management, research grants, scholar evaluation, schematics, slides, posters, venue templates |
| Machine Learning (10) | scikit-learn, PyTorch Lightning, Transformers, SHAP, Stable Baselines3, PufferLib, Torch Geometric, UMAP-learn, aeon, scikit-survival |
| Data Analysis (8) | Polars, Dask, Vaex, NetworkX, GeoPandas, Data Commons, exploratory data analysis, statistical analysis |
| Visualization (6) | Matplotlib, Seaborn, Plotly, scientific visualization, image generation, schematics |
| Research Tools (6) | computational social science, general data science, resource detection, market research, MATLAB, Perplexity search |
| Document Processing (5) | DOCX, PDF, PPTX, XLSX, MarkItDown |
| Databases (4) | OpenAlex, PubMed, bioRxiv, USPTO |
| Quantum Computing (4) | Qiskit, Cirq, PennyLane, QuTiP |
| Materials & Chemistry (4) | Pymatgen, PyMC, PyMOO, FluidSim |
| Physics & Math (3) | Astropy, SymPy, statsmodels |
| Simulation & Engineering (3) | SimPy, Modal, Denario |
Full list with descriptions: interactive catalog
Quick Examples
Literature review with a PRISMA diagram:
Conduct a systematic literature review on CRISPR delivery mechanisms using PubMed and bioRxiv.
Include a PRISMA flow diagram using scientific-schematics.
Publication-quality analysis:
Load my CSV dataset, run exploratory data analysis with statistical tests,
and create publication-quality visualizations with matplotlib and seaborn.
Explainable ML:
Train a classification model on my dataset using scikit-learn,
explain predictions with SHAP, and generate a comprehensive report.
The Quality Gate
What CI enforces today, on every commit (validate.yml):
- SKILL.md present, frontmatter parses,
name/description/license required
- Frontmatter name matches directory name (what the agent loads is what's registered)
- Every relative link resolves to a real file
- Every bundled Python script passes syntax checks and error-level lint
- Registry (
marketplace.json) and disk agree 1:1 — no phantom skills
- Zero vendor-promotion content
- Weekly external-link rot audit (links.yml)
In progress — a deeper editorial audit of all 69 skills across four dimensions: academic rigor, real-world scenario coverage, discoverability of descriptions, and runnability of examples. Findings and fixes land as regular releases.
REVIEW.md is an advisory snapshot of known content and script issues. It is not a CI gate and it is not project policy. The inventory notes in .planning/codebase/CONCERNS.md (2026-04-30) are stale where they disagree with this file or with validate.yml.
Contributing
Factual fixes are the most valuable contribution: wrong API names, outdated signatures, dead links, missing pitfalls. See CONTRIBUTING.md — the validator tells you before CI does:
pip install pyyaml
python scripts/validate_skills.py
License & Attribution
MIT. Original skill content copyright (c) 2025 K-Dense Inc., used and modified under the MIT license; individual skills may carry their own licenses in their SKILL.md frontmatter. This project is not affiliated with or endorsed by K-Dense Inc.
