skillz

My collection of Claude-and-friends skills, agents, and bits.

Skills

• zotero-plugin — Expert guide for building, debugging, and releasing Zotero 7–9 plugins. Covers both the classic bootstrapped stack and the TypeScript zotero-plugin-scaffold / zotero-plugin-toolkit stack.

• gget — Fast CLI/Python queries to ~20 bioinformatics databases (Ensembl, UniProt, NCBI, AlphaFold, Enrichr, ARCHS4, …). Copied verbatim from K-Dense-AI/scientific-agent-skills under their MIT / BSD-2-Clause licensing; see skills/gget/NOTICE.md for provenance.

• gene-set-fetch — Named, provenance-stamped gene sets for human and mouse: transcription factors (Lambert 2018, AnimalTFDB), protein-coding genes (Ensembl biotype, or strict three-way intersection against HGNC/MGI and GENCODE), and set-algebra composites (union, intersection). Fetchers are short and readable on purpose — students should see how the sausage is made.

• ontology-terms — Focused ontology term operations (immediate parents, immediate children, definition + synonyms, free-text search) across four sources: OLS at EBI (general use, version-pinned), Gemma's annotations API (lab alignment / strong search), direct OBO file parsing (max-reproducible), OntoBee SPARQL (flexible). Accepts both GO:0006915 and full OBO URIs; ontology version recorded per call.

• gene-annotations — Bidirectional gene ↔ GO term annotations with default child-propagation: genes_with_annotation <term> and annotations_of_gene <gene>. Two authoritative sources: GOA via QuickGO REST (rich coverage), and GAF (GO Consortium per-species files, version-pinned + reproducible via locally cached go.obo). Symbols resolved transparently. Per-row direct/propagated flag. Term labels filled. Evidence codes preserved. Gemma deliberately NOT a default option.

• gene-statistics — Per-gene meta-statistics for honest interpretation of genomic analyses. v0.1 ships multifunctionality (Gillis & Pavlidis 2011): per-gene MF score over GO BP annotations, ranked 1-N with percentile. Flags genes that appear as top hits in any guilt-by-association / enrichment / DE analysis regardless of the underlying biology. v0.2 will add the Differential Expression Prior (Crow, Pavlidis, Gillis 2019). All evidence codes by default (IEA isn't lower quality). Composes with gene-set-fetch for the protein-coding background.

• enrichment — Gene-set enrichment with MF correction on by default. Two ops: pr_enrichment (Ballouz, Pavlidis & Gillis 2017 PR-AUC method, for scored gene lists) and ora (hypergeometric / Fisher's exact, for hit lists). Generalizable library input: go-bp (default, from GAF + go.obo) or any GMT file (MSigDB, KEGG, custom). MF-correction uses sibling-skill composition with gene-statistics.

• plotting — Publication-quality figure helpers in both Python (matplotlib) and R (ggplot2) with a consistent flat/minimal style. v0.1 ships pavlab_heatmap (expression + correlation + raw), pavlab_scatter (basic, log-scale, categorical/numeric color, hexbin), pavlab_stripchart, and pavlab_density (density, histogram, violin, faceted). N-adaptive alpha/size, viridis colorbars, MF-correction-aware defaults. Comes with a 21-example side-by-side gallery.

• pdf-extract — Reproducible PDF extraction: body text (page-tagged), document metadata (title, authors, DOI, year from XMP/docinfo/first-page heuristics), and annotations (highlights, sticky notes, underlines) with highlighted text resolved from the text stream. Every output ships with a sha256-keyed *_meta.json provenance sidecar. Primary library: pymupdf; text-only fallback: pypdf. No LLM calls.

• provenance-stamp — Write and verify *_meta.json provenance sidecars for any analysis artifact. Records sha256 of the artifact and its inputs, source URL, DOI, upstream version/date, download timestamp (fetched_at), and analysis parameters. Library + CLI; stdlib-only. Compatible with the sidecar format produced by all bioinformatics skills in this repo.

• supplementary-table — Write supplementary tables that name the figure they back. Emits xlsx (with a Provenance sheet) and csv / tsv (with #-prefixed comment header block carrying figure_id, script, source, sha256, stamped_at). One call, both formats, identical hash. Sibling to provenance-stamp for when the artifact IS the table.

• citation-validator — Detect hallucinated bibliography entries. Checks each citation against CrossRef (the DOI registrar): Level 1 — does the DOI exist? (404 = definitively hallucinated). Level 2 — does stated title/year/author match CrossRef's record? (METADATA_MISMATCH catches real DOIs attached to fake papers). Accepts BibTeX and plain text. Level 3 (v0.2): relevance check — does the paper's abstract support the claim at the point of citation? Designed for biolit integration.

• methods-section — Draft a Methods paragraph from an analysis directory. Scans lockfiles (requirements*.txt, renv.lock, conda environment.yaml) for tools and versions, and *.meta.json sidecars for data sources, DOIs, download dates, and parameters. Emits structured JSON; Claude drafts the paragraph. Composes with provenance-stamp.

Installing a skill

Each skill is a directory under skills/ containing a SKILL.md. To install one in Claude Code, either drop the folder into your skills directory or package it into a .skill file with the skill-creator tool and drag it into the app.

Platform note

These skills were developed and tested on macOS. The Python and R helpers aim to be portable — paths go through pathlib / tempfile, text files are opened as UTF-8, and credentials fall back from environment variables to the macOS Keychain — but Linux and Windows are not part of routine CI, so the odd shell-ism or default-encoding wart may still need tweaking. File an issue (or a PR) if something doesn't run for you on another OS.