Turn Your AI Assistant Into a Bioinformatics Expert — For Free

If you're a computational biologist, a bioinformatician, a wet-lab researcher who occasionally runs Python, or anyone who uses an AI assistant to help with life-sciences analysis — SciAgent-Skills is the free install that gives your assistant 199 domain-specific skills in one shot.

It runs inside Claude Code, Codex, Cursor, or Windsurf. And there's a specific, measurable number attached to it: on a public bioinformatics benchmark, adding these skills to Claude Code raised its accuracy from 65% to 92% on the same underlying model. No fine-tuning, no custom model — just structured scientific knowledge the AI can reach for.

(If you don't use one of these AI assistants yet, Claude Code is the easiest way in. And if you don't work in life sciences at all, this specific install isn't for you — but the pattern applies: any professional domain benefits from a skill library that gives the AI real defaults instead of guesses.)

At a glance

What it actually is

A library of 199 scientific skills packaged as plain markdown files. When your AI assistant loads them, it knows the specific tools, parameters, and pitfalls of a huge slice of computational biology — Scanpy for single-cell, DESeq2 for differential expression, RDKit for cheminformatics, AutoDock Vina for docking, PyMC for Bayesian modeling, and about 194 more.

The concrete result: you can now say to your AI assistant "perform differential expression analysis on this RNA-seq count matrix" and it produces a working analysis. Before, it would either guess the tool wrong, guess the parameters wrong, or dutifully write code that compiles but produces nonsense on real data. After, it reaches for the right tool with the right defaults and knows where the common traps are.

The 92% benchmark number comes from BixBench-Verified-50 — a public benchmark run by Future House on real bioinformatics tasks. Claude Code alone scored 65%. Claude Code plus these skills scored 92%. That's a 27-percentage-point jump from the same underlying AI, entirely from the added domain knowledge.

What you get

199 skills grouped into 11 domains — not a general "AI for science" tool, a specific stack of runnable analyses:

• Genomics & bioinformatics (65 skills) — Scanpy, BioPython, pysam, gget, KEGG, PubMed, scvi-tools, Bakta, Roary
• 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

Each skill is a single markdown file with runnable code examples, key parameters, troubleshooting notes, and best practices. There's nothing to install for the skills themselves — the AI just reads the markdown, understands the workflow, and writes the analysis code fresh for your data.

The 21 scientific writing skills are the sleeper hit. They cover manuscript writing, peer review, LaTeX posters, slide decks, and figure guides — meaning the same install that helps you run the analysis also helps you write up the results.

Why this matters for the life-sciences researcher

Three things distinguish SciAgent-Skills from just asking your AI assistant "how do I run scRNA-seq analysis?"

First, the parameter defaults are tuned by someone who has actually run these analyses. Every scientific tool has a small number of choices that make the difference between a valid result and a plausible-looking artifact. Which normalization? Which multiple-testing correction? Which quality-control threshold? The skill files encode the sensible defaults so your AI assistant reaches for them without you having to remember. This is what "192 percentage points from 65 to 92" is actually measuring.

Second, the coverage is wide enough to survive real research. Real projects rarely fit inside one tool. A single-cell analysis needs Scanpy plus scvi-tools plus a way to talk to Reactome plus a downstream statistical test plus a visualization plus a manuscript figure. Fifteen skills fire in sequence and the AI holds the threads. Cherry-picking three skills would give you a demo; 199 skills give you a workflow.

Third, it's genuinely free, genuinely open-source, and genuinely researcher-attributable. The CC BY-4.0 license means you can use these skills for commercial biotech work, teach with them, or fork them for internal tooling — as long as you credit the authors. That's not the license every OSS project ships with, and for a research group considering whether to build their own stack on top, this matters.

How it compares to Bioconductor, commercial platforms, and building your own

Named alternatives across the "how does a computational biologist actually work" landscape:

• Bioconductor — the classic R-based free ecosystem for bioinformatics. Still the standard for many analyses. Completely different mode — Bioconductor is packages you install and call from R directly; SciAgent-Skills is knowledge your AI assistant uses to write the analysis code for you. Complementary, not competitive — SciAgent-Skills often reaches for Bioconductor packages under the hood.
• Nextflow / Snakemake pipelines — the standard way to reproducibly chain bioinformatics tools. Different job. Pipelines are for production, at-scale runs; SciAgent-Skills is for exploration and one-off analyses where the AI is figuring out the right workflow.
• Commercial platforms (Seven Bridges, DNAnexus, Illumina DRAGEN) — the enterprise SaaS route. Point-and-click UIs, hosted compute, high per-run costs, vendor lock-in. SciAgent-Skills runs on your laptop or your own compute, no per-run fees, no lock-in — at the cost of you having to bring the compute.
• Building your own agent skills manually — genuinely viable for a lab with one specific niche, but you'd be starting from scratch. SciAgent-Skills is 199 skills someone already wrote and benchmarked to 92% accuracy. Start here, fork what you need, add your own on top.
• OmicsHorizon — the web platform from the same authors, powered by these skills. If you want the capability without any install at all, this is the zero-setup version.

Verdict

Install — for anyone doing real computational biology on an AI assistant.

The 92% BixBench result is the tell. A 27-point accuracy jump on the same underlying AI, from a free markdown install with no fine-tuning, is not a small thing. If you're the researcher who has been quietly disappointed by asking an AI to help with your Scanpy pipeline and getting subtly-wrong code, this fixes exactly that class of problem.

The one caveat is scope: this is a scientific skill library, not a general research assistant. If your work is legal research, market research, or academic writing outside the life sciences, you want a different install. But if any of RNA-seq, single-cell, drug discovery, proteomics, structural biology, biostatistics, or scientific manuscript-writing is in your regular workflow, install this today.

This post is what we learned from reading the project. A hands-on install-review — running SciAgent-Skills against a real single-cell dataset and reporting where it lifts and where it stumbles — is queued as a follow-up.

For your agent to install this

You can skip this section — it's the install details your AI agent needs. If you're using Claude Code or Cowork, say "install SciAgent-Skills for me" and the agent will read the commands below.

System / OS

• macOS, Linux, or Windows — skills are plain markdown files, no OS-specific requirements
• Git — for cloning the repository
• Python 3.12+ — only needed if you want to run the project's own validation scripts; not required for using the skills
• The skills themselves have no runtime dependencies. Your AI assistant reads the markdown and writes fresh analysis code for your data

API keys — required for any use

None. The skills are markdown files that your AI assistant reads directly. Whatever model provider you already use for your AI assistant runs the analyses — no new key, no new account.

Install commands

Claude Code / Cowork (recommended, persistent install):

/plugin marketplace add jaechang-hits/SciAgent-Skills
/plugin install sciagent-skills

Skills become available as /sciagent-skills:<skill-name> — for example /sciagent-skills:scanpy-scrna-seq. Or just describe the task in plain English and the agent will find the relevant skill automatically.

Session-only load (Claude Code):

git clone https://github.com/jaechang-hits/SciAgent-Skills.git
claude --plugin-dir /path/to/SciAgent-Skills

Codex CLI:

cd your-project
git clone https://github.com/jaechang-hits/SciAgent-Skills.git .sciagent-skills
cp .sciagent-skills/integration-templates/AGENTS.md ./AGENTS.md

Codex reads AGENTS.md at the project root automatically.

Cursor:

cd your-project
git clone https://github.com/jaechang-hits/SciAgent-Skills.git .sciagent-skills
mkdir -p .cursor/rules
cp .sciagent-skills/integration-templates/cursor-rules.md .cursor/rules/sciagent-skills.md

Windsurf and other markdown-compatible agents follow the same "clone + register the skill directory" pattern — see the project's integration-templates/ directory for the specific setup files.

First-run usage

Once installed, just describe the analysis in plain English:

"Perform differential expression analysis on this RNA-seq count matrix"

"Dock this ligand against the PDB structure 3ERT"

"Run single-cell clustering on the attached h5ad file"

The agent picks the relevant skill(s) from the 199-file library, writes the analysis code, and runs it against your data. The canonical routing map — which skill covers which analysis — lives in registry.yaml at the repo root.