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 expert workflow skill: crisp 10-step sequence, fixed A–L output contract, and exemplary progressive disclosure with all 11 referenced modules verified on disk. The main weakness is token efficiency — the same rules are restated across the distinctions list, Core Function, Execution, Hard Rules, and the 'should not' sections — and the absence of any worked example or final output-verification step.
Suggestions
Consolidate redundant rule statements: the "must distinguish between" list, Core Function "should not" list, Hard Rules, and "What This Skill Should Not Do" repeat the same prohibitions (post-baseline predictors, univariable sufficiency, assumed external validation, unvalidated cutoffs) — state each once in Hard Rules and reference it, cutting 30–40% of the body.
Merge the Execution Steps 1–10 with the Mandatory Output Structure A–L (or make one a compact map of the other) since both currently encode the same section-to-reference-module mapping three times (Reference Module Integration, Execution, and Output Structure).
Add a final output-verification checkpoint (e.g., 'Before delivering, confirm all sections A–L are present, every assumption-dependent item is labeled, and one lead protocol version is named') and optionally a short worked example of one filled-in section to make the output contract fully concrete.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Almost every line carries domain-specific design constraints (no space is wasted teaching Claude what biomarkers or survival analysis are), but the ~420-line body restates the same rules repeatedly: the "must distinguish between" list, Core Function (10 items), Execution Steps 1–10, Hard Rules, and "What This Skill Should Not Do" each restate overlapping prohibitions (e.g., post-baseline predictor bias and unverified external validation appear in 3–4 places), and the Execution steps largely mirror the A–L Mandatory Output Structure. This is 'mostly efficient but could be tightened' rather than a 2, since the redundancy is rule-restatement, not explanation of known concepts — but consolidation could cut 30–40% of tokens. | 3 / 5 |
Actionability | For an instruction-only skill the guidance is concrete: every Execution step specifies exactly what to "State" (e.g., Step 3 lists primary endpoint, time origin, follow-up horizon, censoring logic), the Mandatory Output Structure fixes sections A–L with per-section content contracts, and the Clarification Rule bounds questions to "2–6 concise, high-yield follow-up questions". It misses a 5 because some directives remain abstract ("Lead with one coherent main line") and there is no worked example (e.g., a short filled-in sample of sections A–L for one cohort) to anchor the expected output shape. | 4 / 5 |
Workflow Clarity | The 10-step execution sequence is explicit, each step is mapped to a specific reference module and to output sections, and there are genuine checkpoints: input validation with a redirect-and-stop script, a clarification gate before locking the design, and audit steps (Step 8 leakage/overfitting review, Step 9 translation-readiness check). It falls short of 5 because there is no final verify-your-output step (e.g., 'confirm every A–L section is present and assumption-dependent items are labeled before delivering'), so the feedback loop stops at design-time audits rather than output validation. | 4 / 5 |
Progressive Disclosure | The Reference Module Integration section explicitly signals all 11 reference files with one-level-deep paths, assigns each a precise 'use when' condition tied to specific sections/steps, and all 11 files verified present in references/. The body holds the workflow while the rules live in the bundle files — matching the 'clear overview with well-signaled one-level-deep references' anchor; the guard clause ("If any output section is generated without using its corresponding reference module... incomplete") reinforces navigation. | 5 / 5 |
Total | 16 / 20 Passed |