Content
56%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 tool-specific sections are excellent — dense with parameter gotchas, return-shape handling, fallbacks, and a concrete synthesis template — but they are buried in a monolithic 305-line file where roughly a third of the content is general medical knowledge (diagnostic-test heuristics, MCQ test-taking strategy) that Claude already knows and that sits outside the skill's stated purpose. Splitting the file and cutting the generic clinical teaching would sharply improve token efficiency.
Suggestions
Move the Clinical MCQ Reasoning Framework, Diagnostic Test Selection Reasoning, Lab Test Interpretation, and Surgical Decision Making sections into separate reference files (e.g., references/diagnostic-reasoning.md, references/mcq-framework.md) and reference them from a short overview, keeping SKILL.md focused on guideline search and retrieval.
Delete or drastically compress general clinical knowledge Claude already knows (SnNOut/SpPIn, likelihood-ratio thresholds, Brown-Séquard localization, TLICS dimensions, Holliday-Segar formula) — keep only the ToolUniverse-specific workflow facts that cannot be inferred.
Add one fully worked end-to-end example (search across 3 sources with actual calls, parse the returned shapes, then full-text retrieval) so the search-to-synthesis workflow is copy-paste executable, and embed the publication-date/source-verification checks as explicit steps in that sequence.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Roughly a third of the body (~110 of 305 lines: Diagnostic Test Selection Reasoning, Lab Test Interpretation, Surgical Decision Making, and the Clinical MCQ Reasoning Framework) restates general clinical knowledge Claude already knows — SnNOut/SpPIn, likelihood ratios, Brown-Séquard level localization, TLICS dimensions, Holliday-Segar fluids, pre-test probability — rather than ToolUniverse-specific guidance. These are several unnecessary explanation sections unrelated to the skill's core tool usage, matching the 'noticeably verbose' anchor rather than the 'mostly efficient with some unnecessary explanation' one. | 2 / 5 |
Actionability | The tool guidance is highly concrete: exact parameter names and gotchas ("`SIGN_search_guidelines` | `query` (NOT `q`)"), return-shape handling ("use `r.get('data', [])`"), a numbered CPIC workflow with real arguments (`CPIC_get_gene_info(genesymbol='CYP2D6')`), and specific fallbacks ("broaden terms (`'pharmacologic approaches'` not `'metformin first-line'`)"). It stops short of a 5 because there is no fully worked end-to-end example (e.g., a complete search-then-retrieve call sequence with parsed output) tying the pieces together. | 4 / 5 |
Workflow Clarity | The Search Strategy gives a clear 3-step sequence (narrow search across 3+ sources, broaden, retrieve full text) and the Fallback Strategy provides explicit error-recovery loops ("NICE returns empty -> try TRIP or EuropePMC", "ADA returns 0 results -> broaden terms"). Not a 5 because verification checkpoints (e.g., "Always check publication date", "verify the original source") are stated as principles rather than embedded as explicit steps in the sequence. | 4 / 5 |
Progressive Disclosure | The file has good internal structure (clear headers, tables, a template) but is a 305-line monolith with no bundle files and no external references at all. Large self-contained topics — the Clinical MCQ Reasoning Framework, Diagnostic Test Selection, and the per-society tool reference — clearly belong in separate reference files, matching the anchor 'some structure but content that should be separate is inline'. | 3 / 5 |
Total | 13 / 20 Passed |