Content
48%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 delivers excellent domain-specific reasoning (ancestry transferability, winner's curse, PGS-first strategy) and a clear step-0–6 workflow, but it is held back by heavy educational padding Claude doesn't need, pseudocode for its central calculation steps, and a monolithic structure with no reference files. Trimming Key Concepts/Applications into a reference file and making the score-calculation code executable would lift the two weakest dimensions.
Suggestions
Cut or move the "Key Concepts" section (GWAS, LD, effect sizes, population stratification explanations) to a reference file — Claude already knows this genetics material, and it is the largest source of wasted context tokens.
Replace the pseudocode calls (`build_polygenic_risk_score`, `calculate_personal_prs`, `interpret_prs_percentile`) with executable code or the actual ToolUniverse tool invocations, including where weights and genotypes come from.
Offload the PGS Catalog/LDSC tool descriptions and the Disclaimer/Regulatory sections into a `references/` file, keeping SKILL.md as a lean overview with clearly signaled one-level-deep references.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Noticeably verbose with several padded sections: "Key Concepts" re-teaches what GWAS, LD, effect sizes, and odds ratios are — genomics knowledge Claude already has — and "Applications", "Personal Genomics", and the multi-paragraph Disclaimer add further general-knowledge padding. Not 1 because the tool-specific guidance (PGS Catalog params, LDSC usage) is genuinely non-obvious and earns its tokens. | 2 / 5 |
Actionability | The PGS Catalog and LDSC tool guidance is concrete (named tools with parameters and what to read from each response), but the core workflow code blocks are pseudocode — `build_polygenic_risk_score(...)`, `calculate_personal_prs(prs_weights=prs, genotypes=my_genotypes, ...)`, `interpret_prs_percentile(result)` — with undefined functions, no imports, and placeholder variables, so they cannot be executed as written. Not 4 because the central calculation steps lack executable specifics. | 3 / 5 |
Workflow Clarity | Steps 0–6 are clearly sequenced, with a strong step-0 gate (check PGS Catalog first, with explicit fallback conditions for when raw-GWAS building is justified) and QC checkpoints in step 4 (MAF, missingness, HWE, strand ambiguity). Not 5 because there is no explicit validation of the computed score against a reference distribution or any verify-results feedback loop. | 4 / 5 |
Progressive Disclosure | The single ~300-line file is well-headed but inlines content that belongs in separate reference files — educational background (Key Concepts), a detailed PGS/LDSC tool catalog, and regulatory/disclaimer material — and no references/ or scripts/ bundle files exist to offload them. Not 2 because the internal structure (section headers, workflow separation) is solid and navigable. | 3 / 5 |
Total | 12 / 20 Passed |