Content
42%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 is well-organized with genuinely valuable skill-specific guidance (compute-don't-describe discipline, look-up-don't-guess rules, an honesty rule about meta-analysis data requirements, and a concrete tool inventory). However, it is padded with extensive textbook statistical content Claude already knows, contains no executable code despite mandating computation, and references a 'python_implementation.py' that is missing from the bundle.
Suggestions
Move the 'Statistical Methods', 'Study Quality Assessment', and 'Limitations & Caveats' sections (I² formula, fixed/random-effects, heterogeneity sources, quality tiers) into a single reference file, keeping only skill-specific interpretation rules inline — this would cut roughly 170 lines of knowledge Claude already has.
Add at least one executable Python snippet showing the actual meta-analysis workflow (retrieving associations via gwas_get_associations_for_snp, parsing beta/CI from Catalog fields, pooling with inverse variance), since the skill mandates 'write and run Python code' but demonstrates none.
Either ship the referenced 'python_implementation.py' in a scripts/ directory and link it clearly, or remove the reference — as written it points to a file that does not exist in the bundle.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Roughly 170 lines of the body ('Statistical Methods', 'Sources of Heterogeneity', 'Study Quality Assessment', 'Limitations & Caveats') restate textbook meta-analysis and genetics knowledge Claude already knows — e.g. the I² formula, fixed- vs random-effects explanation, and winner's curse explained twice — making it noticeably verbose beyond the score-3 'some unnecessary explanation' anchor. It is not score 1 because genuinely skill-specific guidance ('COMPUTE, DON'T DESCRIBE', 'LOOK UP DON'T GUESS', the honesty rule) is interwoven. | 2 / 5 |
Actionability | Concrete elements exist — specific tool names ('gwas_get_associations_for_snp', 'OpenTargets_get_variant_credible_sets') and exact thresholds (p < 5×10⁻⁸, I² cutoffs) — but there is zero executable code even though the skill instructs 'write and run Python code via Bash', and workflow steps like 'Calculate combined effect size and p-value' lack the concrete commands to execute them, matching the score-3 anchor (concrete but incomplete, missing key executable detail). | 3 / 5 |
Workflow Clarity | Each use case lists a rough sequence (e.g. 'Get top hits from discovery study → Check for presence and significance → ... Calculate replication rate') with stated outcomes, matching anchor 3's 'steps listed but validation gaps; checkpoints missing or implicit'. It does not reach anchor 4 because no workflow includes an explicit validate/verify checkpoint or error-recovery loop, and the analysis steps are high-level rather than executable. | 3 / 5 |
Progressive Disclosure | The body has clear section headers, but large blocks of reference-style material (statistical methods, quality tiers, limitations) are inlined rather than split out, and it cites 'python_implementation.py' which does not exist in any bundle directory (references/, scripts/, assets/ are all absent) — a dangling, unsignaled reference. This fits anchor 3 ('structure present but content that should be separate is inline, references not clearly signaled'), not anchor 2 since section organization is genuinely good, and not anchor 4 because of the broken reference and inlined reference material. | 3 / 5 |
Total | 11 / 20 Passed |