Content
67%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 operational core is excellent — concrete tool invocations, a wrong-vs-correct parameter table, explicit formulas, thresholds, and fallback chains — and the phase workflow is clearly sequenced. However, roughly a quarter of the body re-teaches pharmacology concepts Claude already knows (with duplicated MedDRA guidance and broken 'Reasoning Strategy' numbering), and all four referenced bundle files are missing from the bundle, breaking progressive disclosure entirely.
Suggestions
Ship the four referenced files (SIGNAL_DETECTION.md, REPORT_TEMPLATES.md, CHECKLIST.md, TOOLS_REFERENCE.md) or inline their essential content — currently every reference link in the body is broken, and the skill's checklist, report structure, and example output tables exist nowhere.
Trim the Clinical Reasoning Framework to the skill-specific application notes (which FAERS fields to check, which tools to query, the hypothesis-first sequence) and drop the re-explanations of standard concepts such as on-target/off-target toxicity, Type A/B reactions, the Naranjo algorithm, and the Weber effect.
Deduplicate the MedDRA British-spelling guidance, which appears as a full table in both the 'Tool Parameter Reference' section and Phase 2, and fix the heading numbering (two sections are both titled 'Reasoning Strategy 5').
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Mostly efficient — the tool-parameter tables, phase steps, SIDER frequency/null semantics, and fallback chains are high-value non-obvious information. But the ~90-line 'Clinical Reasoning Framework' re-explains standard pharmacology concepts Claude already knows (on-target/off-target, Type A/B, Naranjo algorithm, dechallenge/rechallenge, Weber effect), and MedDRA spelling guidance is duplicated in both the 'Tool Parameter Reference' table and the Phase 2 table. Fits anchor 3 ('some unnecessary explanation or could be tightened') better than anchor 2 because the majority of the body is skill-specific. | 3 / 5 |
Actionability | Fully concrete guidance: exact tool calls with parameter names ('FAERS_count_reactions_by_drug_event(drug_name=..., limit=50)', 'search_clinical_trials(intervention=..., phase="Phase 3", status="Completed")'), a WRONG-vs-CORRECT parameter table, the PRR formula with variable definitions, explicit thresholds (PRR > 2.0 signal, > 3.0 strong, case count >= 3), the signal score formula with severity weights, evidence tiers, and fallback chains. This matches the anchor-5 'copy-paste ready commands covering the common cases' — the absence of a Python snippet is not penalized per the rubric's instruction-only note, since the formulas and tool invocations are fully specified. | 5 / 5 |
Workflow Clarity | Clear, well-sequenced phase workflow (Phase 0 mechanistic reasoning before tools, through Phase 7 report synthesis) with numbered per-phase steps, mandatory requirements (report-first, citation requirements), fallback chains as tool-failure recovery, and cross-check guidance ('Cross-check against mechanistic prediction', sanity-checking SIDER indications). Falls short of anchor 5 because explicit validate-then-proceed checkpoints are only implied and the phase-by-phase completeness checklist is a dangling reference to a file that does not exist. | 4 / 5 |
Progressive Disclosure | The body links to four bundle files — SIGNAL_DETECTION.md ('detailed disproportionality formulas and example output tables'), REPORT_TEMPLATES.md, CHECKLIST.md, and TOOLS_REFERENCE.md — but no references/, scripts/, or assets/ directories exist in the bundle; every reference is broken. Content the skill depends on (report templates, the verification checklist, example output tables, tool documentation) is neither inline nor present, so the promised one-level-deep disclosure structure fails at navigation time. This matches anchor 2 ('minimal structure... references are buried/broken') better than anchor 3, where references at least resolve. | 2 / 5 |
Total | 14 / 20 Passed |