Content
82%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 high-quality specialist skill body: executable commands, a precise evidence hierarchy with firmness labels, a concrete output contract, and honest limits. The main gaps are minor — some evidence detail duplicated between the body and the reference file, and a papers/ navigation target referenced but missing from the bundle.
Suggestions
Trim the inline hazard-ratio and sample-size detail in rules 1-2 to one-line summaries and defer the numbers to references/evidence-map.md, which already carries them.
Ship references/papers/ (or its INDEX.md) referenced by the "Papers" section, or state explicitly that papers are fetched on demand via paper_fetch.py so the navigation target is not dangling.
Add a brief fallback step for the Run section (what to do if load_check.py errors or returns no activities) to make the workflow's error path explicit.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Every section carries domain-specific evidence content Claude would not know (hazard ratios, cohort sizes, study limits) with no padding of basic concepts. Not 5 because the citation detail in rules 1-2 (e.g. "hazard ratios 1.64 for 10-30 %, 1.52 for 30-100 %, 2.28 for over 100 %") duplicates what is expanded in references/evidence-map.md, and the opening "Paths" blockquote is long for what it conveys. | 4 / 5 |
Actionability | Copy-paste-ready commands with the <skills> placeholder defined up front ("python -X utf8 \"<skills>/rampwise-load-injury/scripts/load_check.py\" --weeks 8" plus a --json variant), a literal handoff block template with exact fields, and a concrete fallback command ("run paper_fetch.py \"<author year>\""). Covers the common cases fully. | 5 / 5 |
Workflow Clarity | Sections follow the natural sequence (run script, interpret flags against the rules, check the evidence before quoting, communicate, emit handoff block) and include checkpoints such as "Before you quote a sample size, effect or limit, open the file" and WARN triage guidance. Not 5 because the sequence is implicit in section order rather than explicit, and there is no recovery step if the script itself fails. | 4 / 5 |
Progressive Disclosure | A well-sectioned ~55-line overview with one-level-deep references, and the referenced references/evidence-map.md and scripts/load_check.py are real bundle files. Not 5 because the body points to "references/papers/" with an "INDEX.md" that is absent from the bundle, and the rules section inlines study detail that also lives in evidence-map.md rather than fully deferring to it. | 4 / 5 |
Total | 17 / 20 Passed |