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.
A well-engineered instruction skill: crisp input validation, a genuinely sequenced workflow with a mandatory pre-output consistency check and feedback loop, and a clean one-level-deep reference bundle. Its main weakness is token efficiency — the Hard Rules section and parts of Step 7 duplicate earlier content almost verbatim — plus the fifth study pattern is never enumerated in the body either.
Suggestions
Cut the redundant 'Hard Rules' recap: rules 1–23 largely restate Steps 3–7 and Sections C.5/G/I/J/D verbatim; keep only genuinely new rules (note the numbering also skips 10 after rule 9) and drop the rest to save ~20 lines of repeated tokens.
Move the detailed A–J output-section specifications and the Step 5 dependency rules into a reference file (e.g., references/output-sections.md), keeping a one-line summary of required sections in the body, to match the disclosure pattern already used for patterns/configs/methods.
Add the concrete ROR computation definition (2×2 contingency table, ROR formula with 95% CI) either inline or explicitly in references/method-library.md, so the analysis steps are directly executable rather than delegated by name.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is written twice: the 23-item 'Hard Rules' section restates Steps 3–7 nearly verbatim (e.g., rules 11–13 repeat Step 4.5's citation rules, rules 15/17 repeat Section G, rule 22 repeats the Section D disclaimer), and Section B/J requirements restate earlier step content. It does not explain concepts Claude already knows, but it contains several padded, purely redundant sections, matching the 2 anchor ('noticeably verbose; several unnecessary... padded sections') better than 3, where redundancy would be incidental rather than structural. | 2 / 5 |
Actionability | For an instruction-only skill the guidance is concrete: an exact out-of-scope redirect script, a pattern-selection table, per-config field requirements, explicit forbidden-claim lists, exact endpoint formulas ('FAERS query + MedDRA PT extraction + comparator + ROR'), and quantified retrieval minimums ('2–4 drug/adverse-event-domain background references'). It falls short of 5 because key analysis mechanics (the ROR/CI computation itself, actual FAERS query construction) are delegated by name without inline anchors, and directives like 'Explain the logic of each decision' remain abstract. | 4 / 5 |
Workflow Clarity | Seven explicitly ordered steps with genuine validation checkpoints and a feedback loop: Step 5 is a mandatory pre-output dependency consistency check with 'If any dependency inconsistency is found, revise the plan before outputting' (validate → fix → re-check), plus checklists (Hard Rules, the six-element Section J review, forbidden-claim gate). This matches the 5 anchor's 'explicit validation steps; feedback loops for error recovery; checklists'. | 5 / 5 |
Progressive Disclosure | The body links 8 reference files, all of which exist, are one level deep, and never reference each other — clearly signaled with '→ Detailed pattern logic: [references/study-patterns.md]' style pointers. However, substantial detail (the full A–J output-section specifications and the Hard Rules recap) is inlined in SKILL.md where this bundle's own pattern would put it in a reference, so organization is good but not optimally split — the 4 anchor. | 4 / 5 |
Total | 15 / 20 Passed |