Content
92%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 strong, expert-level body: fully executable examples with rare version-specific gotchas, a validated workflow with feedback loops, and clean one-level-deep reference structure. The only weakness is mild verbosity - some textbook benchmark material and length that could be trimmed.
Suggestions
Move the effect-size benchmark table (Cohen's d 0.20/0.50/0.80 etc.) and other textbook conventions into references/effect_sizes_and_power.md, keeping only a pointer plus the caveat that benchmarks are conventions - Claude already knows these thresholds.
Tighten the Statistical Integrity section by compressing widely known principles (e.g., 'a p-value does not prove that an effect exists', multiple-comparison basics) into one-line rules, reserving prose for the less-obvious guidance such as the missing-data MCAR/MNAR nuances.
The full APA report templates are illustrative and long; consider keeping one complete exemplar inline and pointing to references/reporting_standards.md for the remaining templates to reduce body length.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense with genuinely non-obvious, high-value content (version-specific gotchas like "Pingouin >= 0.6 column names", "ttest reports absolute Cohen's d", "ArviZ 1.x defaults to 89% intervals"), but includes some textbook material Claude already knows (the Cohen/eta-squared benchmark table "0.20 / 0.50 / 0.80", "a p-value does not prove that an effect exists") and runs long at ~440 lines. It sits between anchors 3 and 4, noticeably above the midpoint: mostly efficient with minor trims available. | 4 / 5 |
Actionability | Every section carries executable, copy-paste-ready code or commands: uv install commands, pg.ttest/pg.anova/sm.OLS/pm.sample blocks with API gotchas inline ("solve_power returns the TOTAL sample size, not n per group"), and fill-in report templates. The common cases (t-test, ANOVA, regression, Bayesian t-test, power) are all covered with complete runnable examples. | 5 / 5 |
Workflow Clarity | The six-step Analysis Workflow is clearly sequenced with an explicit validation checkpoint ("Check assumptions... do not switch tests automatically at a diagnostic p-value threshold"), a dedicated feedback loop for violations ("What to Do When Assumptions Are Violated"), Bayesian convergence gates ("inspect finite R-hat/ESS/MCSE, divergences and posterior predictions before reporting"), and a 7-point "Statistical Integrity" checklist - matching the anchor-5 pattern of sequence plus validation plus error-recovery plus checklist. | 5 / 5 |
Progressive Disclosure | The SKILL.md body is an overview with quick references, and each domain (test selection, assumptions, effect sizes/power, Bayesian, reporting) is split into a real, clearly signaled one-level-deep references/*.md file (all verified to exist), plus a bundled script catalogued with its functions in "Bundled Resources". Navigation is easy and nothing is nested beyond one level. | 5 / 5 |
Total | 19 / 20 Passed |