Content
63%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 delivers excellent tool-specific, non-obvious guidance (exact parameters, MedDRA spellings, formulas, thresholds, fallbacks, data-quality gotchas) in a clear phase sequence, but it is padded with textbook clinical-reasoning exposition and defers critical detail to four referenced files that are not present in the bundle, breaking both actionability and navigation at the mandatory report/checklist steps.
Suggestions
Ship the referenced bundle files (SIGNAL_DETECTION.md, REPORT_TEMPLATES.md, CHECKLIST.md, TOOLS_REFERENCE.md) alongside SKILL.md, or inline their essential content — the mandatory Report-First step and completeness checklist currently point to files that do not exist.
Compress the Clinical Reasoning Framework to the tool-specific 'How to apply this' guidance and move the textbook exposition (Naranjo questions, Type A/B definitions, on-target/off-target basics) into a reference file, cutting roughly 100 lines.
Fix the duplicate 'Reasoning Strategy 5' numbering (Rechallenge and Disproportionality Reasoning are both labeled 5) so the strategy sequence stays navigable.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly efficient — the tool-parameter table, MedDRA spelling table, PRR formula, and SIDER frequency semantics are exactly the non-obvious knowledge Claude lacks — but the ~100-line 'Clinical Reasoning Framework' restates textbook concepts Claude already knows (Naranjo algorithm, Type A/B classification, on-target/off-target basics, dechallenge/rechallenge definitions) before each 'How to apply this' section. This matches 'mostly efficient but includes some unnecessary explanation or could be tightened'; it is not level 4 because the concept exposition is substantial rather than minor, and not level 2 because the tool-specific content genuinely earns its place. | 3 / 5 |
Actionability | Guidance is largely executable: exact tool calls with parameter names ('FAERS_count_reactions_by_drug_event(drug_name=..., limit=50)'), a defined PRR formula with cell meanings, signal thresholds (PRR > 2.0, case count >= 3), a signal-score formula with severity weights, fallback chains, and a SIDER data-quality gotcha with a live example. It falls short of level 5 because the mandatory Report-First step requires report 'section headers' and the PRR calculation details that live in REPORT_TEMPLATES.md and SIGNAL_DETECTION.md — files not present in the skill directory — leaving key details missing. | 4 / 5 |
Workflow Clarity | The Phase 0-7 sequence is clearly laid out with numbered steps per phase, explicit signal thresholds as checkpoints, and a built-in feedback loop ('A signal not predicted mechanistically warrants additional scrutiny'). It is not level 5 because the phase-by-phase verification checklist is deferred to a missing CHECKLIST.md file and there is a duplicate 'Reasoning Strategy 5' heading (Rechallenge and Disproportionality both numbered 5), leaving minor validation and organization gaps. | 4 / 5 |
Progressive Disclosure | References are clearly signaled and one level deep ('See [SIGNAL_DETECTION.md](SIGNAL_DETECTION.md)', '[REPORT_TEMPLATES.md](REPORT_TEMPLATES.md)', '[CHECKLIST.md](CHECKLIST.md)', '[TOOLS_REFERENCE.md](TOOLS_REFERENCE.md)'), but none of these files exist in the skill directory — there is no references/ directory and no sibling files — so the disclosure structure is broken at exactly the points where the most important detail (report structure, checklist, formulas) is deferred. The body is also ~350 lines when the clinical-reasoning exposition belongs in a reference file; this fits 'some structure but could be better organized' rather than level 4, whose 'minor organization gaps' does not cover missing referenced files. | 3 / 5 |
Total | 14 / 20 Passed |