Audited 2026-10-18 · Sources: GitHub repo + README, skill structure, install docs. Tabletop audit — nothing installed.
The card
| Repo | K-Dense-AI/scientific-agent-skills |
| Stars | 52.7k (GitHub API, audit date) |
| License | MIT |
| Born | October 2025 · maintained (recent refresh of all 181 skills) |
| What it is | A library of ~177 validated agent skills for scientific work — literature search, database lookup across 80 sources, document generation, lab-adjacent workflows |
| Install surface | gh skill install K-Dense-AI/scientific-agent-skills (GitHub CLI v2.90+) or the skills.sh installer — works on Claude Code, Codex, Gemini CLI, Antigravity, Cursor |
What problem it solves
Every research workflow an agent does well, somebody had to teach it first. This repo is 177 pre-taught workflows — scientific database retrieval with provenance checks, document pipelines, slide decks. If your operation touches research, citations, or technical content, that's a library you didn't have to write.
Our gate findings
| Gate | Finding |
|---|---|
| Upstream identity | Verified — K-Dense-AI org, MIT, companion BYOK desktop app |
| Engineering hygiene | Best on the list. Every skill carries a SKILL.md with purpose/workflow/version metadata; skills with scripts have their own test suite and a declared requirements file. That's how a skills library should be built. |
| Honest provenance | They removed vendored Anthropic document skills when licensing got ambiguous — that's what a maintainer does when they take provenance seriously |
| Instruction surface | Skills are prompt files — they shape your agent's behavior. "Validated" means the maintainers tested them, not that a security review passed. Read a skill before adopting it. |
| Data boundary | Skills that hit live databases/services send queries out by design — know which ones phone home before pointing them at client data. |
Verdict: GO. Low install risk (scripts + text you can read before loading), high engineering quality, MIT. The residual risk is the nature of skills — you're adopting someone's prompt engineering — not anything specific to this repo.
Best for
Research-heavy operators — content teams doing evidence-based writing, clinics checking literature, anyone whose agents currently reinvent scientific workflows on the fly.