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, actionable skill body: concrete tool names, executable code, a clear five-step workflow, and unusually strong honesty/verification rules. The main improvements are trimming already-known Python basics from the cookbook, adding explicit error-recovery loops, and externalizing the reference-style repository catalog.
Suggestions
Trim the Data Loading Cookbook to the non-obvious formats (XPT/sas7bdat, fwf, URL download with format detection, API pagination) and drop the basic loaders (`pd.read_csv`, `pd.read_excel`, `pd.read_stata`) and the generic matplotlib scatter example — Claude already knows these.
Add explicit error-recovery feedback loops to the workflow, e.g., 'if a repository search returns nothing, fall back to the literature strategy in Layer 3' and 'if a download fails or is access-restricted, report the access tier and try the next-ranked candidate'.
Move the Layer-2 per-tool repository details (re3data, DDBJ, Synapse annotations) into a references/ file, keeping SKILL.md as the decision-rule overview with one-line pointers per repository.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense with non-obvious domain knowledge (design-to-question mapping, repository tiers, fitness criteria, honesty rules) and avoids concept-explaining prose. It slips below a 5 because the Data Loading Cookbook includes basics Claude already knows — `pd.read_csv("data.csv")`, `pd.read_excel`, and the generic matplotlib scatter block — that could be trimmed to the non-obvious formats (XPT, fwf, URL format-detection). | 4 / 5 |
Actionability | Fully executable throughout: named MCP tools (`re3data_search_repositories`, `DHSProgram_search_indicators`, `Synapse_search_entities`, `DbGaP_search_studies`), copy-paste-ready Python for downloads/pagination/analysis, and concrete heuristics ('~800 subjects at 80% power', '~10 events per predictor'). This matches the anchor for fully executable, copy-paste-ready guidance covering common cases. | 5 / 5 |
Workflow Clarity | A clear five-step sequence (Understand → Search → Evaluate → Download/Analyze → Report) with a Step-3 checklist and real checkpoints ('download a small sample first and inspect', 'verify the codebook', 'Always run the code and report actual numbers'). It misses a 5 because explicit error-recovery feedback loops (e.g., what to do when a download fails or returns nothing) are absent. | 4 / 5 |
Progressive Disclosure | The single file is well-sectioned with clear headers and every section is plausibly needed at decision time, and no bundle files exist to organize. It stays at 4 rather than 5 because the ~45-line loader cookbook and the per-tool Layer-2 repository catalog (e.g., the long re3data paragraph) are reference-style content that could live in a references/ file for a skill this broad. | 4 / 5 |
Total | 17 / 20 Passed |