Content
60%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, sequenced workflow with genuine one-level-deep reference modules, weakened by substantial rule redundancy — the same prohibitions are restated in four separate sections, and several sections duplicate each other's content. Consolidating the rule lists and trimming the overlapping sections would materially cut token cost without losing guidance. The main functional gap is the absence of a worked example or final validation checkpoint.
Suggestions
Consolidate 'Core Function should not', 'Hard Rules', 'What This Skill Should Not Do', and 'Quality Standard' into a single Hard Rules section — they currently restate the same fabrication, post-baseline-confounder, and causal-overreach prohibitions four times.
Replace the 'Sample Triggers' section with the Input Validation examples (they overlap heavily), or merge both into one 'When to use / when to redirect' section.
Move detail-heavy inline material (the cohort-family lists and per-rule explanations) into the corresponding reference files, which are currently much thinner than the main body, and add one worked example of a filled-in output section (e.g., a sample Section D eligibility table).
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body restates the same constraints in four overlapping places: 'This skill should not' (Core Function), 'Hard Rules', 'What This Skill Should Not Do', and 'Quality Standard' all repeat the no-fabrication, no-post-baseline-confounder, and no-causal-overreach rules; 'Sample Triggers' duplicates the Input Validation examples; and 'Supported Cohort Study Families' overlaps 'Cohort Design Selection Logic'. This is noticeably verbose with several padded/redundant sections, fitting level 2 rather than level 3, where verbosity would be occasional rather than systemic (roughly half the body could be cut without losing information). It is not level 1 because there is no padding that explains concepts Claude already knows — the material is domain rules, just repeated. | 2 / 5 |
Actionability | As an instruction-only skill, code absence is not penalized, and the guidance is directly executable: Steps 1–11 each enumerate exactly what to state ('index date / time-zero', 'censoring rules', 'primary endpoint... operational definition'), with concrete formatting rules (tables in sections D/E/F/G/H/J, necessary/recommended/optional split in G). It is not level 5 because there is no worked example of any output section (e.g., a sample eligibility table or one filled-in endpoint definition), and not level 3 because the steps are complete enough to execute without inventing structure. | 4 / 5 |
Workflow Clarity | Steps 1–11 are clearly sequenced, each bound to its reference module and to the A–L output sections, with an input-validation gate (redirect-and-stop for out-of-scope) and a bounded clarification rule (2–6 questions). It is not level 5 because there is no explicit final validation checkpoint (e.g., re-audit the drafted protocol against the Hard Rules before delivering) — the bias/feasibility audits exist as content steps but not as a verify-then-fix loop. This is a design skill, not a destructive/batch operation, so the level-3 cap does not apply. | 4 / 5 |
Progressive Disclosure | All 12 referenced files exist in references/, contain real one-level-deep content (spot-checked), and are explicitly mapped to specific sections/steps in the 'Reference Module Integration' list and per-step callouts. It is not level 5 because the split is unbalanced: SKILL.md carries ~23KB of inline rule detail while each reference is only ~0.5KB, so material (e.g., the long Hard Rules and cohort-family lists) is inlined that could live in the reference files; it is above level 3 because navigation is easy and every reference is clearly signaled with its use case. | 4 / 5 |
Total | 14 / 20 Passed |