Content
60%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 well-structured methodology skill with a clear step sequence, useful lookup tables, and genuine validation thinking, but it under-delivers on execution: database queries are named rather than shown, and the causal-inference primer spends tokens re-teaching concepts Claude already knows. The three tables and evidence-grading scheme are its strongest assets.
Suggestions
Add one executable example of an actual ToolUniverse database query (e.g., a short Python snippet retrieving OpenTargets or Reactome evidence for one finding), since find_tools discovery strings alone don't show how to perform the skill's core operation.
Trim the causal-inference primer in Step 3 (confounder/mediator/collider definitions, basic MR logic, 'association is not causation') down to the domain-specific prescriptions (F-statistic > 10, MR-Egger intercept check, triangulation recipe) that Claude would not produce on its own.
Close the validation loop explicitly in Step 4: state that when genetic evidence contradicts the analysis direction (or the negative control shows an association), the finding must be downgraded or the contradiction resolved before the Step 5 report is written.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Mostly efficient with dense tables, but Step 3 re-explains concepts Claude already knows ('Statistical association is not causation', definitions of confounders/mediators/colliders, basic MR logic), which is more than the minor trimming of the anchor above. | 3 / 5 |
Actionability | Concrete elements exist (find_tools query strings, F-statistic > 10 threshold, report template), but the skill's central operation is left abstract ('Query GWAS Catalog or OpenTargets', 'Search PubMed/EuropePMC') with no executable example of an actual database query — key details are missing. | 3 / 5 |
Workflow Clarity | A clear five-step sequence with an explicit falsification stage ('Before reporting a finding as robust, attempt to falsify it') and structured output templates, but there is no explicit fix-and-recheck loop after a failed cross-validation, keeping it below the full validation-loop anchor. | 4 / 5 |
Progressive Disclosure | No bundle files exist and the single-file body is cleanly sectioned with no nested or buried references, fitting the good-structure anchor; it does not reach the top anchor because the causal-reasoning frameworks could reasonably live in a reference file and no well-signaled references exist. | 4 / 5 |
Total | 14 / 20 Passed |