SciAgent-Skills
Benchmark: 92.0% on BixBench-Verified-50
OmicsHorizon, powered by SciAgent-Skills, achieved 92.0% accuracy on the BixBench-Verified-50 benchmark — outperforming all other systems compared. Notably, Claude Code (Opus 4.6) without skills scores 65.3%, but jumps to 92.0% simply by equipping it with these domain-specific skills (+26.7%p).
196 ready-to-use scientific skills for AI coding agents — covering genomics, proteomics, drug discovery, biostatistics, scientific computing, and scientific writing.
Each skill is a self-contained SKILL.md file with runnable code examples, key parameters, troubleshooting guides, and best practices. Designed for Claude Code, but compatible with any agent that reads markdown skill files.
What's Inside
| Category | Skills | Examples |
|---|---|---|
| Genomics & Bioinformatics | 63 | Scanpy, BioPython, pysam, gget, KEGG, PubMed, scvi-tools |
| Structural Biology & Drug Discovery | 26 | RDKit, AutoDock Vina, ChEMBL, PDB, DeepChem, datamol |
| Scientific Computing | 24 | Polars, Dask, NetworkX, SymPy, UMAP, PyG, Zarr, SimPy |
| Cell Biology | 15 | pydicom, histolab, FlowIO |
| Biostatistics | 12 | scikit-learn, statsmodels, PyMC, SHAP, survival analysis |
| Scientific Writing | 21 | Manuscript writing, peer review, LaTeX posters, slides, figure guides |
| Systems Biology & Multi-omics | 11 | COBRApy, LaminDB, Reactome, STRING |
| Proteomics & Protein Engineering | 10 | ESM, UniProt, PyOpenMS, matchms, HMDB |
| Lab Automation | 6 | Opentrons, Benchling |
| Data Visualization | 5 | Plotly, Seaborn |
| Molecular Biology | 3 | CRISPR sgRNA design, gene expression, cloning |
Skill types: 72 toolkits, 53 database connectors, 36 guides, 35 pipelines
Installation
Prerequisites
- Claude Code CLI installed
- Git
- Python 3.12+ (only needed if you want to run validation scripts)
Step 1: Clone the Repository
git clone https://github.com/jaechang-hits/SciAgent-Skills.git
cd SciAgent-Skills
Step 2: Choose Your Setup Method
Method A: Claude Code Plugin (Recommended)
Load SciAgent-Skills as a Claude Code plugin for the current session:
claude --plugin-dir /path/to/SciAgent-Skills
To verify the plugin loaded, run /plugin inside Claude Code and check that sciagent-skills appears in the Installed tab.
Skills become available as /sciagent-skills:<skill-name>:
/sciagent-skills:scanpy-scrna-seq
/sciagent-skills:rdkit-cheminformatics
/sciagent-skills:pymc-bayesian-modeling
Or just describe your task — the agent finds the relevant skill automatically:
"Perform differential expression analysis on this RNA-seq count matrix"
Persistent installation — to load the plugin automatically in every session, use the plugin install command inside Claude Code:
/plugin marketplace add jaechang-hits/SciAgent-Skills
/plugin install sciagent-skills
Method B: Project-Level Integration
Clone into your project directory so Claude Code picks up skills via CLAUDE.md:
cd your-project
git clone https://github.com/jaechang-hits/SciAgent-Skills.git .sciagent-skills
Add to your project's CLAUDE.md:
## Scientific Skills
Reference skills in `.sciagent-skills/skills/` for domain-specific analysis.
Registry: `.sciagent-skills/registry.yaml`
Step 3: Install Dependencies
cd SciAgent-Skills
pixi install
Pixi handles the Python environment and all required packages. If you don't have pixi installed:
curl -fsSL https://pixi.sh/install.sh | bash
How Skills Work
Each skill follows a structured template:
skills/<category>/<skill-name>/
SKILL.md # Main skill file (300-550 lines)
references/ # Optional deep-dive reference files
assets/ # Optional templates, configs
A SKILL.md contains:
- Frontmatter — name, description, license (for agent discovery)
- Overview & When to Use — what the tool does and when to reach for it
- Prerequisites — packages, data, environment setup
- Quick Start — minimal copy-paste example
- Workflow / Core API — step-by-step pipeline or module-by-module API guide
- Key Parameters — tunable settings with defaults and ranges
- Common Recipes — self-contained snippets for common tasks
- Troubleshooting — problem/cause/solution table
The agent reads only the description field during planning. Full skill content is loaded on demand when relevant.