Content
56%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
The body has strong domain-specific value — the Reasoning Strategy, PGS Catalog tool guidance with gotchas, and the gated step-0 workflow are excellent — but it is padded with textbook genetics Claude already knows, redundant repetition between the Reasoning Strategy and Limitations sections, and an inlined conceptual primer that belongs in a reference file. The undefined build/calculate/interpret functions in the workflow are the main actionability gap.
Suggestions
Cut or collapse the 'Key Concepts' section (GWAS/OR/beta/LD explanations) — Claude already knows this material; keep only the skill-specific claims (ancestry transferability magnitude, winner's curse policy) that are already well-covered in Reasoning Strategy.
Merge the 'Limitations and Considerations' section into the Reasoning Strategy to eliminate the duplicated winner's-curse, ancestry-bias, and not-diagnostic points, and move the PGS Catalog per-tool documentation into a references/ file (e.g., references/pgs-catalog-tools.md) so SKILL.md stays an overview.
Resolve the provenance of build_polygenic_risk_score, calculate_personal_prs, and interpret_prs_percentile — either list them as actual tools with their parameters in Data Sources or replace them with the real tool calls the raw-GWAS fallback path should use.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Noticeably verbose with several padded sections: the 'Key Concepts' section (~44 lines) explains textbook genetics Claude already knows (what GWAS are, beta/OR definitions with worked examples, 'PRS forms a bell curve', LD basics), 'Limitations' largely restates the Reasoning Strategy section (winner's curse and ancestry bias each appear twice), and time-sensitive detail ('FDA does not currently regulate PRS (as of 2024)') sits outside any old-patterns section. This exceeds 'some unnecessary explanation' — several distinct sections could be cut or folded together. | 2 / 5 |
Actionability | Mostly executable guidance: concrete tool calls with real parameter names (PGSCatalog_search_traits(query=...), gwas_get_associations_for_trait, gnomad_get_gene_constraints), specific QC thresholds (MAF < 0.01, HWE p < 1e-6, strand-ambiguous A/T and G/C removal), and concrete percentile cutoffs. The gap is that build_polygenic_risk_score, calculate_personal_prs, and interpret_prs_percentile are invoked in workflow steps 2/5/6 but never defined as tools in the Data Sources list, leaving those snippets without clear provenance. | 4 / 5 |
Workflow Clarity | Steps 0–6 are clearly sequenced, and step 0 acts as an explicit gate ('check this FIRST') with well-defined fallback conditions ('Only fall back... when PGS Catalog has no score... or when every existing score was developed... in an ancestry that doesn't match'). Checkpoints are present (ancestry_distribution check before trusting a score, performance metrics by ancestry, SNP QC filters), but minor validation gaps remain — e.g., no handling for a trait with no GWAS Catalog associations, and held-out validation appears only as a research best practice rather than a workflow step. | 4 / 5 |
Progressive Disclosure | No bundle files exist and the skill is a ~300-line monolithic SKILL.md. Internal section structure is good, but content that should be separate is inlined — the Key Concepts primer, the detailed per-tool PGS Catalog documentation, and the Applications survey would each fit a reference file. The under-50-line simple-skill exception does not apply. | 3 / 5 |
Total | 13 / 20 Passed |