163 skill files give AI agents the domain knowledge raw LLMs can't improvise

An AI coding agent knows Python. What it doesn't know is which of the 78 public databases to query for a given task, how to avoid the 0-based vs 1-based coordinate trap that silently corrupts a genomic analysis, or when a PK/PD model output is a research artifact versus something that requires a qualified pharmacokineticist's sign-off. Scientific Agent Skills is a collection of 163 SKILL.md files that supply exactly that layer — curated documentation, version-pinned examples, and explicit safety boundaries — for agents running on any host that supports the open Agent Skills standard.

The scope is genuinely broad: bioinformatics, cheminformatics, proteomics, clinical pharmacology, medical imaging, quantum computing, geospatial science, laboratory automation, regulatory evidence preparation. A single database-lookup skill covers 78 databases rather than 78 separate skills — the README explains this was a deliberate consolidation that cut always-on context cost by 13.9x while holding routing accuracy across five models. That kind of design reasoning appears throughout: the waypoint-bio skill exists because an unconverted MetaPhlAn table retains only 3% of abundance mass and still returns a valid embedding, silently.

The safety boundaries are unusually explicit. Clinical reporting skills produce visibly marked drafts from verified source-fact manifests only. The PK/PD skill covers non-compartmental analysis, population PK, bioequivalence, and first-in-human dose estimation using stdlib plus numpy/scipy — no proprietary estimation software — but the output is a research artifact, not a dosing decision. Regulatory skills draft ISO and ICH evidence-preparation artifacts and are careful to note that laboratories are accredited rather than certified, and that ISO 15189 accreditation does not satisfy CLIA. These distinctions matter and the README states them plainly.

Security gets real treatment. Skills are scanned weekly using Cisco AI Defense Skill Scanner, with a full rescan at least every 30 days. The README explicitly warns against installing the full collection indiscriminately now that community contributions have grown, and recommends reading each SKILL.md before installing. Every skill that ships scripts/ must have a test suite; a structural CI contract checks frontmatter conformance, link resolution, and --help behavior on every pull request.

Installation works via npx skills add, GitHub CLI gh skill install with version pinning, or manual clone. The companion K-Dense BYOK project wraps these skills into a local desktop research workspace supporting 40-plus models.

The honest limitation: 163 skills add up to substantial standing context, and the README says so directly — install a topical subset rather than everything. Individual skill licenses may also differ from the repository's MIT license, so per-skill review is the user's responsibility.