nexus-scientific

name	nexus-scientific
description	Entry point for 149 vendored scientific research skills (K-Dense Scientific Agent Skills). Use for ANY task in bioinformatics/genomics (scanpy, biopython, pysam, phylogenetics, RNA-seq, variant calling), cheminformatics/drug discovery (rdkit, deepchem, datamol, docking, ADMET, medchem), proteomics/mass-spec, structural biology (esm, diffdock, molecular dynamics), clinical research and medical imaging (pydicom, pathology, EHR, trials, treatment plans), materials science and computational chemistry, physics/astronomy (astropy, qiskit, sympy, quantum), statistics and scientific ML (statsmodels, pymc, scikit-learn, survival, power analysis, forecasting, SHAP), geospatial science, and the research workflow itself — literature review, paper/database lookup, hypothesis generation, experimental design, peer review, grants, scientific writing, figures, posters, slides, and schematics. Also fires on any mention of a specific scientific library or database in that space. Routes to exactly ONE sub-skill; never loads all 149.
license	MIT (upstream K-Dense-AI/scientific-agent-skills)

nexus-scientific

The estate's scientific research library. 149 curated skills covering scientific libraries, 100+ databases, and research workflows — vendored from K-Dense-AI/scientific-agent-skills (MIT) and distributed to every machine through the skills-library hub.

This file is a router. It is the only thing in the global skill roster; the 149 sub-skills live under library/ and are checked out on demand, per the Library doctrine in ~/.claude/CLAUDE.md §6.

Routing protocol

1. Find the skill. Grep the index — never read it whole:

   grep -i -E 'rna-seq|single.cell' "$HOME/.claude/skills/nexus-scientific/library/INDEX.md"
   

   library/INDEX.md is a | dir | what it does | table over all 149. Match on the library name, the database name, the method, or the deliverable.

2. Check out ONE. Read library/<dir>/SKILL.md and follow it exactly. It may point at its own references/, scripts/, or examples/ — read those only when that skill tells you to.

3. Check in. When it has served its purpose, stop re-reading it. Do not summarise the sub-skill back to the user; do the work it describes.

Never read more than one sub-skill speculatively, and never cat library/ wholesale — that is 23 MB and will bury the session.

If the grep returns nothing, say so and fall back to normal tool use. A miss is a real answer; do not force an unrelated skill onto the task.

Domain quick-map

One hop for the common asks. Anything not listed here → grep INDEX.md.

If the task is about	Start at
Single-cell / scRNA-seq	scanpy, anndata, scvi-tools, scvelo, cellxgene-census
Bulk RNA-seq / DE	bulk-rnaseq, pydeseq2, pathway-enrichment
Sequences, alignments, variants	biopython, pysam, scikit-bio, phylogenetics, gget
Cheminformatics / molecules	rdkit, datamol, molfeat, medchem, deepchem
Docking, binding, protein structure	diffdock, esm, molecular-dynamics, torchdrug
Proteomics / mass spec	pyopenms, matchms
Clinical / EHR / trials	pyhealth, clinical-decision-support, clinical-reports, treatment-plans
Medical imaging / pathology	pydicom, histolab, pathml, imaging-data-commons
Stats, Bayesian, survival, power	statsmodels, pymc, scikit-survival, statistical-power, statistical-analysis
ML / interpretability / forecasting	scikit-learn, pytorch-lightning, shap, timesfm-forecasting, transformers
Physics, quantum, symbolic maths	astropy, qiskit, pennylane, qutip, sympy, fluidsim

Estate rules that still bind

The sub-skills are third-party. Estate doctrine wins on conflict:

• Python runs through uv. Sub-skills quote bare pip install / python x.py. Use uv run / uv pip install so the environment stays reproducible, per [[feedback_full_green_before_handover]]. Never install into system Python.
• Public-facing words route through nexus-copywriter, then brand-guardian. A sub-skill writing a paper, abstract, poster, or report does not bypass that gate if the output is client- or public-facing.
• Claims need real sources. literature-review / paper-lookup output is research substrate, not verified fact — the evidence standard in [[feedback_no_false_reporting]] applies. Cite what you actually retrieved.
• Secrets. Several sub-skills (benchling-integration, dnanexus-integration, modal, latchbio-integration, ginkgo-cloud-lab, opentrons-integration, labarchive-integration) want API keys. Pull them from 1Password (op), never hardcode, and check skills/library/connections.md first.
• Credit/compute spend. Sub-skills that hit paid compute (modal, optimize-for-gpu) or wet-lab / cloud-lab APIs (adaptyv, ginkgo-cloud-lab, opentrons-integration, tamarind) are real-money or real-world actions. Gate them per autonomy-ladder — confirm before spending or before anything that moves physical lab hardware.

Provenance and refresh

• Upstream: K-Dense-AI/scientific-agent-skills, MIT — see library/UPSTREAM-LICENSE.md.

• Pinned commit: library/UPSTREAM-COMMIT.

• To pull a newer upstream snapshot:

  sh "$HOME/.claude/skills/nexus-scientific/refresh.sh"
  

  It re-vendors library/, rewrites UPSTREAM-COMMIT, and regenerates INDEX.md. Review the diff, then commit to skills-library so all machines get it on the next sh ~/.claude/bootstrap.sh.

Sub-skills are vendored verbatim. Do not edit anything under library/ — a refresh overwrites it. Estate-specific behaviour belongs in this file.