Content
66%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-structured with a genuinely rigorous five-step methodology and strong built-in validation via the cross-validation step. It is dragged down by textbook re-explanations of epidemiological concepts Claude already knows and by the absence of any executable code example despite the skill's own 'compute, don't describe' mandate.
Suggestions
Trim Step 3's background explanations of Mendelian randomization, DAG confounder/mediator/collider definitions, and 'association is not causation' down to only the applied decision rules (instrument strength F>10, pleiotropy checks, what to adjust for).
Add one worked, copy-paste-ready example of the full pipeline the skill mandates — e.g., a find_tools() query, retrieving UniProt/Reactome annotations for a short gene list, and the pandas merge that produces the evidence summary table.
Move the evidence-grading table and the causal-reasoning frameworks (DAG, triangulation, MR, mediation) into a reference file linked from SKILL.md so the main file stays a lean overview.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Mostly efficient — tables are dense and every section carries decision rules — but Step 3 re-explains concepts Claude already knows ("genetic variants (instruments) are assigned at conception, so they are not confounded", definitions of confounders/mediators/colliders, "Statistical association is not causation") that could be tightened to just the applied heuristics. This fits 'mostly efficient but includes some unnecessary explanation or could be tightened'; it is well above the padded verbosity of the 2 anchor. | 3 / 5 |
Actionability | Concrete elements exist — exact `find_tools("...")` query strings, a 4-tier evidence grading table, and a filled-in report template — but no executable code despite the body mandating "write and run Python code via Bash... Don't describe what you would do -- execute it". The tool-discovery snippets are effectively pseudocode hints with no worked example of a query→pandas→report pipeline, matching 'some concrete guidance but incomplete; pseudocode instead of executable code'. | 3 / 5 |
Workflow Clarity | A clear five-step sequence (map findings to questions → multi-database evidence → causal reasoning → cross-validation → structured reporting) with an explicit validation checkpoint: "Before reporting a finding as robust, attempt to falsify it" plus five falsification checks. Feedback loops for error recovery are present ("If MR contradicts observational data, suspect confounding"; "there is a contradiction to resolve"), matching the top anchor. Not below 5 because nothing here is a destructive/batch operation requiring additional checkpoints. | 5 / 5 |
Progressive Disclosure | No bundle files exist (references/, scripts/, assets/ are absent), and the ~125-line body is well-sectioned with clear headers, tables, and no buried or nested references. This fits 'good structure; most content appropriately placed; minor organization gaps' (4) rather than 5, since the evidence-grading tables and causal-reasoning frameworks are reference-grade material that could be split out to keep SKILL.md a leaner overview. | 4 / 5 |
Total | 15 / 20 Passed |