The Problem

Life-science researchers and bioinformatics engineers increasingly rely on AI coding agents like Claude Code, but these agents lack domain-specific knowledge. A generic model will fumble RNA-seq pipelines, misinterpret BAM files, or hallucinate API parameters for tools like Scanpy or RDKit. The result is slow, error-prone analysis that requires heavy human oversight.

What This Does

SciAgent-Skills is a library of 197 self-contained SKILL.md files, each teaching an AI agent one bioinformatics task—from scanpy-scrna-seq to rdkit-chemdraw-cdxml. Each skill bundles runnable code, key parameters, and troubleshooting guidance. The README reports a 92.0% accuracy on BixBench-Verified-50, up from 65.3% for a baseline Claude Code (Opus 4.6). The repo also includes scaffolding and validation tooling under .claude/skills/sciagent-skill-creator/ and scripts/ to author and test new skills.

How It Is Wired

Execution starts at main in .claude/skills/sciagent-skill-creator/scripts/scaffold.py (line 224), which reaches 26 functions and is the only entry point. It validates skill metadata (validate_name, validate_tags, validate_category) and renders new skill files via render_skill_md, which touches the filesystem by reading a template. The fail function is the hub—called from 18 places across the codebase—so any change to error handling ripples widely. The scripts/validate_registry.py parse_frontmatter (6 call sites) and scripts/blind_knowledge_test.py invoke_claude (2 call sites, makes an outbound network call for benchmarking) are the other key touchpoints.

The internal call graph shows 84 resolved edges with no cycles, so refactoring is low-risk. The tests/test_skill_quality.py (43 functions) is the largest test file, exercising skill content quality; scripts/validate_registry.py is called from 3 other files, making it the most reusable validation module.

How To Use It

Setup: Clone and install dependencies via pixi (the pixi.toml and pixi.lock files indicate a pixi-managed environment):

git clone https://github.com/moses-y/SciAgent-Skills
cd SciAgent-Skills
pixi install

Configuration: No environment variables are required for basic use. The blind_knowledge_test.py script reads a .env file for AWS Bedrock credentials, but that's only for running the benchmark, not for using the skills.

Running it: For Claude Code, the skills load automatically from the .claude/skills/ directory. To validate your own skill, run python scripts/validate_registry.py. To scaffold a new one, run python .claude/skills/sciagent-skill-creator/scripts/scaffold.py. The README documents no other launch commands.

Real-World Use

A researcher with a batch of raw FASTQ files can prompt Claude Code with "Run differential expression using DESeq2" and the agent will pull the deseq2-differential-expression skill, execute the correct pipeline, and interpret the output. The integration-templates/ folder shows how to wire these skills into Cursor or Windsurf, extending the same capability beyond Claude Code.

Code Health & Issues

Static analysis (deterministic, not opinion) found 7 findings: 1 high, 6 medium. The high-severity issue is [cognitive_load] deep nesting (max indentation depth 9) in build_reaction_scheme.py, blind_knowledge_test.py, and validate_registry.py—control flow is hard to follow; fix with guard clauses. Medium issues include [resilience] broad exception handling in check_scheme.py (bare except swallows errors) and [cognitive_load] high branching density in check_scheme.py and test_registry.py (68 branch points over 207 lines).

Beyond that, the CI audit flags two high-severity SDLC issues: GitHub Actions are pinned to mutable tags (prefix-dev/setup-pixi@v0.8.0) rather than commit SHAs, and the workflow never runs the test suite despite 3 test files existing. The validate.yml also lacks least-privilege permissions and job timeouts. No secrets or license issues were found.

The Bottom Line

This is a genuinely useful, well-organized skill library with a strong benchmark result and clean internal architecture. The main risks are operational—unpinned CI dependencies and a test suite that isn't wired into the pipeline. For any team running bioinformatics through AI agents, this is worth adopting; just fix the CI before trusting the green check.