Scientific Agent Skills

What It Does

K-Dense Scientific Agent Skills is a collection of 139 SKILL.md files that teach AI coding agents how to work with scientific Python libraries, public databases, and lab automation platforms. The skills span bioinformatics, cheminformatics, clinical research, proteomics, medical imaging, materials science, physics, engineering, geospatial analysis, and scientific communication. Each skill is a structured markdown document with code examples, best practices, and reference material that an agent can read on demand. The repo also includes 352 helper scripts and 697 reference files. It follows the open Agent Skills standard and installs via npx skills add or gh skill install, targeting Cursor, Claude Code, Codex, and Gemini CLI.

The Good

Massive domain coverage with real depth. 139 skills organized across 17 scientific categories. The database-lookup skill alone covers 78 public databases (PubChem, ChEMBL, COSMIC, ClinicalTrials.gov, NASA, NIST, USGS, SEC EDGAR) with individual reference files for each one. The rdkit skill is 780 lines with code examples for molecular I/O, descriptors, fingerprints, reactions, and 3D coordinate generation. The scanpy skill walks through a complete single-cell RNA-seq pipeline from data loading to visualization.

Consistent frontmatter and structure across skills. Every SKILL.md has YAML frontmatter with name, description, license, and metadata (skill-author). All 139 skills have descriptions. The internal structure is uniform: Overview, When to Use This Skill, code examples organized by capability, and references to supporting files. This consistency makes the pack predictable for both agents and human reviewers.

Real install tooling and CI/CD. The repo uses npx skills add (the standard Agent Skills installer), gh skill install with version pinning, and has GitHub Actions for PR skill scanning and security analysis. The scan_skills.py script runs behavioral, trigger, and LLM-based security analysis via Cisco AI Defense Skill Scanner. The security scan results are published in SECURITY.md with severity badges and remediation steps. This is more security infrastructure than most skill packs provide.

Progressive disclosure for large skills. The database-lookup skill has a routing table that maps user intent to specific database references, so the agent reads only the relevant reference file instead of loading all 78. Skills like exa-search separate routing logic from detailed instructions in separate reference files. This respects the agent's context window.

Transparent security reporting. The SECURITY.md file, dated 2026-05-18, lists all 856 findings across 138 skills: 68 CRITICAL, 18 HIGH, with only 31 of 138 skills marked safe. The report is honest about what it found and includes specific remediation steps per skill. This transparency is rare in skill packs.

The Bad

68 CRITICAL security findings and 107 unsafe skills. The security scan flags 20 skills as CRITICAL, including clinical-decision-support, treatment-plans, research-grants, literature-review, and scientific-writing. These findings include behavioral patterns like environment variable access combined with network calls, unpinned package installations, and skill names that impersonate legitimate open source libraries (e.g., zarr-python mimicking the zarr-developers project). While some findings are inherent to the skill type (teaching an agent to run code and install packages), the volume is concerning. The README itself warns: "Do not install everything at once."

Quality variance across skills. SKILL.md sizes range from 56 lines (ginkgo-cloud-lab) to 1594 lines (latex-posters). Some skills are thin vendor brochure pages (ginkgo-cloud-lab reads like a pricing page for Ginkgo Bioworks). Others, like consciousness-council (150 lines), are creative writing exercises rather than scientific tools. The 29 community-contributed skills come from multiple authors with varying depth standards.

Proprietary Anthropic content mixed into an MIT repo. The pdf and pptx skills include Anthropic's proprietary LICENSE.txt files with restrictions against extracting, reproducing, or creating derivative works. These are Anthropic's terms, not MIT, creating a license mismatch in a repo that declares itself MIT-licensed.

Heavy clone size. A shallow clone is 74 MB, driven by the 697 reference files, 352 scripts, and supporting assets. For users who only need a few skills, cloning the entire repo or running npx skills add pulls everything. There is no per-skill install via npx (though gh skill install supports targeting individual skills).

Smoke Test Results

We cloned the repo and ran structural validation on the host (macOS, aarch64).

Structural validation

$ git clone --depth 1 https://github.com/K-Dense-AI/scientific-agent-skills /tmp/scientific-skills-r3
Cloning into '/tmp/scientific-skills-r3'...
✅ clone succeeded

$ find /tmp/scientific-skills-r3 -name "SKILL.md" | wc -l
139
✅ 139 SKILL.md files found (README claims 138, one was added recently)

$ du -sh /tmp/scientific-skills-r3
74M
✅ total repo size under 100 MB

$ wc -l /tmp/scientific-skills-r3/scientific-skills/*/SKILL.md | tail -1
62426 total
✅ 62,426 total lines across all SKILL.md files

$ find /tmp/scientific-skills-r3 -name "references" -type d | wc -l
127
✅ 127 skills have reference directories

$ find /tmp/scientific-skills-r3 -path "*/references/*" -type f | wc -l
697
✅ 697 reference files found

$ find /tmp/scientific-skills-r3 -path "*/scripts/*" -type f | wc -l
352
✅ 352 script files found

$ for f in /tmp/scientific-skills-r3/scientific-skills/*/SKILL.md; do awk '/^---/{n++; next} n==1 && /^description:/{found=1; exit} END{if(!found) print FILENAME}' "$f"; done
✅ all 139 skills have description fields in frontmatter

$ grep -c "skill-author:" /tmp/scientific-skills-r3/scientific-skills/*/SKILL.md
139
✅ all 139 skills have skill-author metadata

$ grep -h "skill-author:" /tmp/scientific-skills-r3/scientific-skills/*/SKILL.md | sort | uniq -c | sort -rn
105 skill-author: K-Dense Inc.
✅ 105 of 139 authored by K-Dense Inc. (76%)

$ cat /tmp/scientific-skills-r3/pyproject.toml | grep version
version = "2.39.0"
✅ version is 2.39.0

$ git log --format="%cs" -1
2026-05-21
✅ last commit date matches README claim of active maintenance

$ head -8 /tmp/scientific-skills-r3/scientific-skills/pdf/LICENSE.txt | head -2
© 2025 Anthropic, PBC. All rights reserved.
❌ Anthropic proprietary license found in MIT-licensed repo (pdf, pptx skills)

$ cat /tmp/scientific-skills-r3/SECURITY.md | grep "CRITICAL" | wc -l
29
❌ 29 lines referencing CRITICAL severity findings

$ cat /tmp/scientific-skills-r3/SECURITY.md | grep "Safe skills:" | head -1
**Safe skills:** 107/138
❌ only 31 of 138 skills passed the security scan as safe

Pass rate: 10 of 12. Two structural failures: proprietary license files in an MIT repo, and the security scan's low safe-skill ratio.

Setup Walkthrough

1. Install via npx (requires Node.js):

 npx skills add K-Dense-AI/scientific-agent-skills

This copies the skill directories to your agent's skills folder.

1. Or install a specific skill via GitHub CLI:

 gh skill install K-Dense-AI/scientific-agent-skills scanpy

1. Or clone directly for inspection:

 git clone --depth 1 https://github.com/K-Dense-AI/scientific-agent-skills

The clone is 74 MB. Each skill lives in scientific-skills//SKILL.md.

1. Prerequisites: Python 3.11+ (3.12+ recommended), uv package manager, and a supported agent (Cursor, Claude Code, Codex, Gemini CLI). Individual skills have their own Python dependencies listed in the SKILL.md.

1. After install, your agent discovers skills automatically. You can also invoke a skill by mentioning its name in your prompt.

Post-install note: The README explicitly recommends installing only the skills you need, not the entire collection, due to security concerns with community contributions.

Alternatives

1. anthropics/skills -- Anthropic's official skill pack with fewer skills but stricter quality control and no mixed-license issues. Better for general-purpose agent augmentation.
2. google/skills -- Google's agent skill collection. Smaller and more Google-platform-oriented, but well-maintained and security-reviewed.
3. agentskills/agentskills -- The Agent Skills specification and reference implementation. Provides the standard that K-Dense follows, plus a curated set of foundational skills.