Statistical Genomics Foundry
Design docs design record

bioSkills

Our nearest neighbor on the repository-of-skills axis — its framing, strengths, and where it sits on our values.

github.com/GPTomics/bioSkills — "a set of SKILLS.md for doing bioinformatics with agents like Claude Code," by GPTomics. ~952★, "Version 3.0" (Feb 2026), hundreds of skills across 63 categories. Our nearest neighbor on the "repository of skills" axis. Verified 2026-06-26 against the repo, skill_writing_reference.md, and sample SKILL.md files.

Its framing

bioSkills presents itself as a broad, practitioner-grade library of agent skills for bioinformatics — sequence I/O, alignment, variant calling, scRNA-seq, ChIP/ATAC, Hi-C, phylogenetics, pop-gen, metagenomics, proteomics — plus a workflows category of end-to-end pipelines. The unit is the hand-authored SKILL.md: it is the deliverable and the source of truth. The project's signature is CLI-rigor discipline — a mandatory ## Version Compatibility block, a primary_tool, introspection commands, and a "Per-Tool Failure Modes" template, all enforced by a 19-point validation lint. Young, fast-moving, genuinely popular.

Strengths relative to our approach

On several axes bioSkills is ahead of the Foundry today:

  • CLI-integration rigor is its signature, and it is excellent. Verbatim ## Version Compatibility (pinned versions + introspection: pip show, packageVersion(), --version), an instruction to adapt to the installed API on ImportError/AttributeError rather than retry blindly, example scripts tagged with a verify-if-version-differs reference, and a structured failure-mode table. More disciplined than typical skill repos about the agent–tool version-mismatch problem. This is the closest external match to a value we hold.
  • Embedded statistical wisdom. Domain skills carry genuine correctness content ("'Tajima's D < −2 means a sweep' is wrong by construction"; "|iHS| > 2 is an empirical convention, NOT a calibrated -value"), confound callouts, and citations. High-quality content — distinct from an automated referee.
  • Coverage and momentum — hundreds of skills, 63 categories, active.
  • Genuinely multi-runtime with a documented format-conversion step.
  • Structured relationship fields (depends_on, qc_checkpoints, Related Skills) — more formal than prose-only repos.

Where it sits on our values

Value bioSkills
Produces skills Yes — hundreds of hand-authored SKILL.md. This is exactly our packaging target; bioSkills is our nearest neighbor here, not a contrast.
Progressive disclosure Yes, standard — the Agent Skills shape (frontmatter description → SKILL.md body → references//scripts/), with a documented conversion step. No layer beneath the skill, though — disclosure is within the artifact, not from an authoritative source.
Traceability Partial — skills carry citations and a tool-version header, but the SKILL.md is the source ("you are responsible for every line"). No upstream inspectable source-of-truth, no per-artifact provenance lineage; Related Skills links are plain-text paths, format-validated for presence, not machine-resolved. A reader can see what a skill cites, but cannot trace a line back to a governed source.
CLI instructions derived from CLI docs Strong (by discipline) — mandatory version-compatibility + introspection + adapt-on-mismatch, enforced by lint. Derived through an authoring contract, not an automated pipeline from CLI documentation, but the value is squarely honored.
Portability High — installers for Claude, Codex, Antigravity (Gemini), OpenCode, OpenClaw, plus a documented conversion to the Agent Skills standard. Multi-target.
Human scrutiny Yes, at the file — skills read as human-practitioner prose; readable, correctable, and contributable by hand (PRs). But there is no navigable, progressive-disclosure learning surface — it is a directory of files, not a site a human reads to learn the domain.
Knowledge-base backed No — the skills are the source of truth; there is no inspectable knowledge base beneath them.

Where it diverges from the Foundry

In bioSkills the SKILL.md files are the source of truth: hand-authored, carrying citations and a tool-version header, but with no governed source beneath them. The Foundry inverts that stack — an inspectable abstract source-of-truth (Molds) is cast into portable skill artifacts that carry per-artifact provenance, with machine-resolved cross-references and a navigable knowledge base a human reads to learn. Its embedded statistical guidance is excellent content; the Foundry's distinction is an empirical referee that gates a result, not better prose. And bioSkills' CLI discipline — version pins, introspection, adapt-on-mismatch — is a genuine strength the Foundry shares rather than contrasts. The short version: bioSkills is the artifact layer done well; the Foundry builds the layer beneath it.

Sources