Content
71%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-structured, highly actionable phase-by-phase pipeline with exact tool contracts, decision thresholds, and edge-case fallbacks. Its main weakness is progressive disclosure: it is a single ~290-line monolith with no bundle files, inlining per-tool API detail and interpretation tables that should live in referenced files.
Suggestions
Move per-tool input/output contracts and the FAERS interpretation tables into references/ files (e.g., references/tools.md, references/faers-interpretation.md), keeping SKILL.md as a lean overview with one-level-deep pointers.
Add a short executable example in 'COMPUTE, DON'T DESCRIBE' (e.g., a sample FAERS disproportionality call followed by a pandas/scipy analysis) so the compute mandate is copy-paste ready.
Either integrate the data-harmonization guidance (ICD-10/SNOMED, informative missing data) into a specific phase or cut it, and add explicit validation checkpoints such as confirming the resolved SPL/drug identity before running Phase 1 queries.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and operational — tool I/O contracts, parameter gotchas, and thresholds — with no padding on basics Claude already knows. A few sections could be trimmed, e.g., the unanchored data-harmonization paragraph (ICD-10 vs SNOMED, missing lab values) that is not tied to any phase. | 4 / 5 |
Actionability | Every tool lists exact inputs/outputs ('medicinalproduct' not 'drug_name', 'setid' NOT 'set_id', 'limit=1000') plus explicit signal criteria (PRR >= 2.0, lower CI > 1.0, N >= 3). Minor gap: 'COMPUTE, DON'T DESCRIBE' mandates Python analysis but provides no example call or code snippet. | 4 / 5 |
Workflow Clarity | Clear Phase 0-6 sequence with per-phase objectives, tools, workflows, decision criteria, error-handling notes ('NOT_FOUND response for boxed warnings is normal', MedDRA term mismatch), edge cases, and evidence grading. Falls short of a 5 because it lacks explicit feedback loops, e.g., re-validating resolved drug identity before proceeding. | 4 / 5 |
Progressive Disclosure | No bundle files exist (references/, scripts/, assets/ are absent) and the ~290-line body inlines reference-grade material — per-tool API detail, FAERS interpretation tables, signal-credibility criteria — that belongs in separate reference files with pointers from the overview. Well-sectioned, but the split is overdue. | 3 / 5 |
Total | 15 / 20 Passed |