Content
92%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-built skill body: every phase gives concrete commands, file conventions, validation checkpoints, and error-recovery paths, with details correctly pushed to real one-level-deep reference files. The only weakness is mild sentence overload where multiple caveats are compressed into single statements.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | No padding or explanations of concepts Claude already knows; the body is dense with tables, exact commands, and conditional rules. A few rule-packed sentences (e.g. the Phase 4 caveats paragraph and Phase 2's output-directory rules) bundle many conditions into one sentence and could be tightened for clarity, keeping it below a 5. | 4 / 5 |
Actionability | Fully executable guidance: copy-paste commands with exact flags ('node scripts/aggregate-rank.js {输出目录} --out {输出目录}/扫榜聚合.md --sparse 10 --scale short' and the --sample variant), concrete output-directory naming ('扫榜/{YYYYMMDD}/' with '-2' suffix), verbatim author-facing questions, and a complete report template with placeholders. | 5 / 5 |
Workflow Clarity | A clear 5-phase sequence with explicit validation (check the aggregate's '采集情况' for failed titles/collections) and a feedback loop (troubleshoot and re-collect per platform reference; if unfixable, state in the report that conclusions exclude it), plus conditional rules like skipping the '变化' section for single-period samples. | 5 / 5 |
Progressive Disclosure | The body is a well-signaled overview: per-platform commands, URLs, troubleshooting, and analysis dimensions are split into five reference files that all exist and are one level deep (they point only to scripts/, never to each other), while the emotion table and report template that are needed on every run stay inline. | 5 / 5 |
Total | 19 / 20 Passed |