Content
73%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.
The body is an unusually rigorous, actionable judgment-layer skill with an exemplary staged workflow and checkpoint system, but it severely overruns token budget for a SKILL.md: meta-governance material and inlined detail that belongs in reference files make it far longer than runtime execution requires. The weakest dimension is conciseness; structure and disclosure are strong but undercut by the file's monolithic size.
Suggestions
Move the skill-maintenance meta sections (rule-admission procedure, grep verification steps, blind-test protocol) into a separate reference file (e.g. references/skill-governance.md) and keep a one-line pointer in the body, cutting a large block of non-runtime tokens.
Move the dynamic-animation event schema, transfer ledger, and gate 9 measurement/tolerance procedures into a dedicated reference (triggered by the stated conditions), keeping only the trigger criteria and core rules inline.
Trim each war-story to the single most calibrating quote and eliminate full restatements of a rule across stage, gate, and anti-pattern table — the anti-pattern section already declares itself a pointer table, so stages/gates can reference instead of repeating.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | At ~750 lines / 87KB the body is noticeably verbose: it carries skill-maintenance meta sections (rule-admission procedure, grep verification steps, blind-test protocol) that are not runtime guidance, plus repeated restatements of the same rules across stages, the delivery gate, and the anti-pattern table, and long war-story quotations. It does not explain basics Claude already knows, so it stays above the lowest anchor, but several sections are padded or duplicative. | 2 / 5 |
Actionability | For an instruction-only skill the guidance is fully executable: quantified criteria ("top-2 时点占了合计的多少？过半就得在图上让这件事可见"), concrete formulas (`0.05 + 0.5*sqrt(value/max)`), a precise animation event schema (`{ id, track, operation, label, delta, from, to, at, duration }`), and gate items that demand observable artifacts (screenshots, mapping tables, measurement records) rather than self-assessment. | 5 / 5 |
Workflow Clarity | The five-stage sequence (定意图 → 验数据 → 选形 → 编码 → 交付前自检) is explicit, each stage ends in a handoff checklist of required artifacts, and the delivery gate is a DO-CONFIRM checklist with a conditions-based exemption table and error-recovery guidance ("写不出来的那条就是没过"; the systematic-analysis section after being critiqued). Validation checkpoints are everywhere and mechanically checkable. | 5 / 5 |
Progressive Disclosure | The seven reference files all exist, are one level deep, and are surfaced through a well-signaled "参考文件 | 什么时候读" table anchored to stages plus inline pointers — genuinely good structure. However, the SKILL.md itself is not an overview: substantial self-contained detail (the dynamic-animation event ledger, gate 9's measurement tolerance procedures, the Sankey criteria) is inlined where the established pattern would put it in a reference, leaving the main file monolithic. | 4 / 5 |
Total | 16 / 20 Passed |