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.
The ToolUniverse-specific core is excellent — precise parameter quirks, return formats, gotchas, and fallback loops that could only come from live verification. However, the skill is bloated by generic clinical-reasoning material Claude already knows, and a ~300-line monolith with no reference files wastes the progressive-disclosure structure the format supports.
Suggestions
Move the "Clinical MCQ Reasoning Framework" and "Diagnostic Test Selection Reasoning" sections into separate reference files (e.g., references/mcq-reasoning.md, references/diagnostic-reasoning.md) and keep only a one-line pointer plus the trigger conditions in SKILL.md, cutting the body by roughly a third.
Delete or compress explanations of concepts Claude already knows — SnNOut/SpPIn, sensitivity vs specificity basics, pre-test probability, likelihood ratios, Brown-Séquard anatomy, Holliday-Segar fluids — retaining only the tool-specific application of those concepts.
Add one or two worked example sequences showing the parallel Step-1 searches with actual parameter values (e.g., NICE_Clinical_Guidelines_Search(query='heart failure guideline', limit=10) alongside TRIP and EuropePMC calls) to make the opening workflow as executable as the CPIC section already is.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The tool-workflow sections are dense and earn their tokens, but roughly a third of the body restates clinical knowledge Claude already has — SnNOut/SpPIn mnemonics, pre-test probability, likelihood ratios, Brown-Séquard lesion anatomy, Holliday-Segar fluids, TLICS scoring — plus a ~65-line "Clinical MCQ Reasoning Framework" only tangentially related to guideline retrieval. This fits "mostly efficient but includes some unnecessary explanation or could be tightened"; not 2 because the padding is terse reference bullets rather than padded prose and the tool-specific material is genuinely non-obvious. | 3 / 5 |
Actionability | Guidance is copy-paste ready throughout: exact tool names with required parameters ("`SIGN_search_guidelines` — `query` (NOT `q`)"), literal example calls ("CPIC_get_gene_info(genesymbol='CYP2D6')"), PostgREST syntax ("genesymbol='eq.CYP2D6'"), return-shape parsing ("r.get('data', [])"), per-failure fallbacks, and a synthesis template. As an instruction-only skill, the specific examples cover the common cases, matching the anchor-5 description; not 4 because there are no real gaps in executability of the guidance given. | 5 / 5 |
Workflow Clarity | The search strategy is clearly sequenced (start narrow with parallel calls → minimum 3 sources → full-text retrieval) with explicit error-recovery feedback loops ("NICE returns empty -> try TRIP or EuropePMC", "TRIP returns 403/gated PDFs -> note limited access, try alternative sources") and verification habits (check publication date, check returned `status` field). Not 5 because the checkpoints are scattered across separate sections rather than woven into one workflow, and some verification is implicit; not 3 because explicit fallback loops and checklists are present. | 4 / 5 |
Progressive Disclosure | There are no bundle files at all — the entire ~305-line skill is inline. Section headers and tables give it real structure, but the tool catalog, the diagnostic-test selection reasoning, and the MCQ reasoning framework each clearly belong in separate reference files, matching the anchor-3 pattern of "content that should be separate is inline". Not 2 because the in-page organization and navigation via headers is good; not 4 because at this length the lack of any file split is a real structural gap, not a minor one. | 3 / 5 |
Total | 15 / 20 Passed |