Content
62%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 impressively thorough, well-gated roleplay specification — its CHECKPOINT system and three-step agentic protocol are exemplary workflow design. Its weakness is size: at ~765 lines, with duplicated blocks and reference-grade biographical/source material inlined, it reads as a monolithic document rather than a lean overview over the existing references layer.
Suggestions
Move the 人物时间线 table, 智识谱系, 最近动态, and 附录 source lists into a file under references/ (e.g., references/biography.md), keeping only the roleplay-relevant identity summary in SKILL.md, and signal the split files from the relevant sections the way the appendix currently points to references/research/.
Delete the verbatim duplicated blocks in 模型 5 (场景 3 and 场景 4 appear twice back-to-back) and audit other sections for repeated evidence lines — these are pure token cost.
Add a fallback for Step 2's research tool: 'mcp__MiniMax__web_search' will be absent in most environments, so specify a substitute (any available web-search tool) or an explicit skip path so the workflow's 必须使用工具 mandate cannot dead-end.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body runs ~765 lines with noticeable padding: Model 5's 场景 3/场景 4 blocks are duplicated verbatim ("'抵挡汤是绝对唯一处方…'" appears twice back-to-back), the 人物时间线 and 智识谱系 sections carry biographical detail marginal to the clinical roleplay, the 附录 re-lists primary/secondary sources already documented in references/research/, and a 最近动态 (2024-2026) news section adds little operational value. Not score 1: most of the material is genuinely novel domain knowledge (dosage rules, case evidence, expression DNA) that Claude does not already know, not explanation of familiar concepts. Not score 3: the duplication and appendix redundancy go beyond 'some unnecessary explanation'. | 2 / 5 |
Actionability | Guidance is highly concrete and copy-paste ready for an instruction-only skill: the verbatim 免责声明 text, the 三问法则 self-check, CHECKPOINT scripts with exact quoted phrases ("中药和西药不要同时吃，会影响彼此的吸收"), the 分批追问 rounds with the specific questions to ask, the eight-step 回答模板, and the 表达DNA taboo-word list. Not score 5: the research workflow depends on a specific tool ('⚠️ 必须使用工具（mcp__MiniMax__web_search）') that may not exist in most environments with no fallback, and several model sections mix instructive rules with background evidence narrative. | 4 / 5 |
Workflow Clarity | The agentic protocol is a clearly sequenced multi-step process with explicit validation gates: Step 1 problem-classification table routing each question type to an action, Step 2 five research dimensions with search guidance and a defined output hand-off ('研究完成后，先在内部整理事实摘要…然后进入 Step 3'), Step 3 model-selection table plus answer template. The seven CHECKPOINTs act as explicit blocking conditions ('三轮追问完毕、关键四诊仍 < 3 项时不开方') with error-recovery paths (待补充四诊 list), and the 退出机制 defines three exit tiers. This matches the top anchor: clear sequence, explicit validation, feedback loops. | 5 / 5 |
Progressive Disclosure | References exist, are real (verified: references/research/ holds 6 files totaling exactly the claimed 1,933 lines), and are one level deep — but they are mentioned only once, in the appendix ('调研过程详见 references/research/ 目录（6 个文件，共 1,933 行）'), and are not used as a disclosure layer for the body's bulk. Large blocks that clearly belong in separate files are inlined in SKILL.md itself: the 26-row timeline table, 智识谱系, the source appendix, and the full 心智模型/决策启发式 evidence sections. Not score 2: the body is well-sectioned with headers throughout and the reference layer is present and accurate. Not score 4: far too much reference-grade material is inlined rather than split out. | 3 / 5 |
Total | 14 / 20 Passed |