Content
86%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.
An exceptionally well-engineered skill body: copy-paste-ready CLI examples at every step, an explicit capability matrix with fail-closed routing, validation gates in the workflow, and a properly structured one-level-deep reference bundle that was verified to exist. The only meaningful weaknesses are minor token overhead from the dated version-pin and citation sections and the lack of an explicit error-recovery loop in the workflow.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body consistently assumes Claude's competence — no basic concept is explained; every section is dense task-specific policy (capability matrix, I/O contract, 10 EDA rules, output-interpretation caveats). Minor trimmable spots remain: the dated version-pin table and the multi-paragraph citation instructions are time-sensitive content outside any 'old patterns'/'deprecated' section. Fits anchor 4 (efficient, minor instances that could be trimmed) — not anchor 5 given those sections, and clearly above anchor 3 since nothing teaches Claude what it already knows. | 4 / 5 |
Actionability | Every workflow step is backed by copy-paste-ready commands with concrete flags: `python scripts/capability_manifest.py inspect data.csv --root /approved/project --output data.manifest.json`, `eda_analyzer.py ... --max-rows 100000`, `missingness_leakage_audit.py ... --group-column condition`, through `report_scaffold.py --analysis-date 2026-09-30`. The examples cover the common cases across the automated tiers, matching anchor 5 (fully executable, copy-paste ready, common cases covered). | 5 / 5 |
Workflow Clarity | The five-step workflow (confirm authorization → manifest before content → run narrowest tool → add scientific context → create report) is clearly sequenced with explicit validation checkpoints: step 1 stops on unclear authorization, step 2 gates on manifest status ("If status is reference_only, do not run eda_analyzer.py... If unknown, stop"), and the Output interpretation section acts as a result-verification checklist. It falls short of anchor 5 only in lacking an explicit error-recovery feedback loop (what to do when a command fails or output looks anomalous mid-run); it is well above anchor 3's implicit/missing checkpoints. The destructive/batch cap does not apply since the skill is read-only by design. | 4 / 5 |
Progressive Disclosure | Verified against the actual bundle: all six `references/*.md` files, `assets/report_template.md`, and every invoked `scripts/*.py` exist; the body is a true overview (capability matrix, routing, workflow) with a well-signaled one-level-deep reference table plus the instruction "Read the one relevant format reference. Do not load every reference"; reference files link only to external sources, with no nested chains. This matches anchor 5 (clear overview, well-signaled one-level-deep references, easy navigation). | 5 / 5 |
Total | 18 / 20 Passed |