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.
A code-rich, largely executable skill whose statistical decision checkpoints (variance/normality routing, multiple-testing corrections, batch-effect thresholds) are its real value. However, it spends most of its token budget re-teaching standard scipy usage Claude already knows and repeats provider-compatibility marketing, and it keeps everything inline instead of splitting domain use cases into reference files.
Suggestions
Cut or drastically compress the basic-library tutorials (t-test, correlation, non-parametric, normality snippets) down to one-liners or a cheat-sheet table; keep only what Claude would not already know, such as the domain use cases and decision thresholds.
Move the Common Use Cases (differential expression, cluster enrichment, batch effect detection) into reference files (e.g., references/use-cases.md) and keep SKILL.md as a concise overview with clearly signaled pointers.
Remove the repeated 'Works with ALL LLM providers' claims, the boilerplate ✅/❌ print blocks, and unused imports (proportions_ztest, kstest) to tighten the token budget.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The majority of the body (Steps 1-6 and Advanced Features, ~260 lines) re-demonstrates standard scipy/statsmodels calls Claude already knows (ttest_ind, pearsonr, mannwhitneyu, shapiro), plus 'Works with ALL LLM providers' is repeated in the Overview and Technical Notes, ✅/❌ print boilerplate pads every snippet, and the Troubleshooting section addresses trivial issues. This matches anchor 2 ('noticeably verbose; several unnecessary explanations or padded sections') rather than 3, since it is not 'mostly efficient' — well over half the tokens re-teach known library usage. | 2 / 5 |
Actionability | Guidance is mostly executable with concrete, runnable patterns (Levene's test routing to Welch's t-test, multipletests FDR/Bonferroni with p-clipping, Fisher exact enrichment contingency construction, Cohen's d). Anchor 4 rather than 5 because several snippets depend on undefined placeholder inputs (p_values, gene_names, gene_expression_df, data) and imports like proportions_ztest and kstest are never demonstrated. | 4 / 5 |
Workflow Clarity | Steps 1-6 are clearly sequenced with real validation checkpoints: Levene's test routes to Welch's t-test, Shapiro-Wilk routes to non-parametric tests, batch-effect detection has a >10% warning threshold, and Troubleshooting covers NaN cleaning and sample-size minimums. Not 5 because there is no explicit end-to-end error-recovery loop tying the checkpoints into one validated workflow. | 4 / 5 |
Progressive Disclosure | The body has clear section headers, but no bundle files exist at all — the entire ~490-line skill is one monolithic file, and the domain-specific use cases (differential expression, cluster enrichment, batch effect) clearly belong in separate reference files. This matches anchor 3 ('some structure but content that should be separate is inline'). | 3 / 5 |
Total | 13 / 20 Passed |