Content
63%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's greatest strengths are domain-specific operational intelligence — live-verified tool bugs, parameter renames, and a float-underflow filtering pitfall with working code — plus an explicit workflow and concrete output schema. It is dragged down by a monolithic single-file structure with no reference files, textbook genetics explanations Claude doesn't need, and Best Practices code invoking an undefined function.
Suggestions
Move the tool catalog details, Programmatic Access snippets, Limitations, and Citation into reference files (e.g., references/tool-notes.md, references/programmatic-access.md) and keep SKILL.md as a workflow + key-gotchas overview.
Cut concepts Claude already knows from Key Concepts, Best Practices, and Limitations (LD, association vs. causation, sample-size effects), keeping only skill-specific thresholds and decision rules.
Replace or define the undefined discover_gwas_genes() calls in Best Practices with real invocations of the cataloged tools (e.g., gwas_search_associations with the documented filter parameters).
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The operational core is tight and high-value (broken-tool warnings, parameter renames, the pvalue_mantissa/pvalue_exponent underflow gotcha), but several sections explain genetics Claude already knows: "Linkage Disequilibrium: Lead SNP may tag the true causal variant in a nearby gene", "Association ≠ Causation", "Larger studies detect more associations", "p-values alone don't indicate effect size", plus textbook definitions in Key Concepts. Sections like Citation, Support, Data Sources coverage stats, and the Use Cases question list add padding that could be trimmed. This matches 'mostly efficient but includes some unnecessary explanation or could be tightened' rather than 4, where only minor instances would remain. | 3 / 5 |
Actionability | Mostly executable: tool-by-tool catalog with gotchas, copy-paste-ready code ("tu.tools.gwas_get_variants_for_trait(trait=\"type 2 diabetes\", size=200, sort=\"p_value\", direction=\"asc\")", the genome_wide_significant filter function, REST download snippets), and a concrete output schema. It is not 5 because the Best Practices section calls "discover_gwas_genes(...)" — a function defined nowhere in the skill or its (nonexistent) scripts — so those blocks are not executable as written, and the REST snippets are fragments rather than complete examples. | 4 / 5 |
Workflow Clarity | The 5-step workflow (Trait Search → SNP Aggregation → Gene Mapping → Evidence Ranking → Annotation) is clearly sequenced, each step maps to named tools, and confidence-level rules plus fallback guidance ("This tool returns errors for most queries. Use gwas_search_associations(query=<trait>) as a reliable alternative") act as checkpoints. Not 5 because there are no explicit validation/error-recovery checkpoints embedded in the workflow itself (e.g., what to do when a trait search returns nothing) and the steps are not tied to their tools inline; not 3 because the sequence is unambiguous and practical fallbacks for known-broken tools are documented. | 4 / 5 |
Progressive Disclosure | The single 305-line file is well-sectioned with clear headers, and the one external pointer ("See tooluniverse-data-wrangling skill for pagination, bulk download, and format parsing patterns") is clearly signaled and one level deep. However, no bundle files exist and everything is inlined in one monolithic document — the 17-tool catalog, Programmatic Access details, Limitations, and Citation material would naturally live in reference files, and the body functions as a full manual rather than an overview pointing to detail. This fits 'some structure but could be better organized; content that should be separate is inline'. | 3 / 5 |
Total | 14 / 20 Passed |