Content
53%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 skill body is a clear, well-structured overview that correctly offloads detail to a real references/README.md and a scripts/ bundle, with good input/output and constraint documentation. Its main weaknesses are actionability (the generation steps are abstract prose with no executable code) and workflow validation (no concrete validate→fix→retry loop despite apply-mode write-back).
Suggestions
Add one concrete executable example for the core flow (e.g. the actual command or python call to skill_controller.py with preview/apply), so the 7-step process becomes copy-paste actionable rather than prose.
Turn the '校验' section into an explicit validation loop with a concrete script command (e.g. `python scripts/format_guard.py ...`) and a defined pass/fail criterion before apply-mode write-back, to satisfy the destructive/batch feedback-loop expectation.
Tighten or relocate the '公共硬约束' block to a shared reference so the body keeps only the skill-specific constraints, improving token efficiency.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly efficient and well-sectioned, but the large '公共硬约束' block and repeated config.yaml/path-convention material partially restate cross-skill policy Claude could derive, and several statements are abstract rather than tight instruction; it sits above 2 (no heavy padding or beginner explanations) but below 4. | 3 / 5 |
Actionability | It gives some concrete pointers (config.yaml paths, {project_path}/.complete_example/<run_id>/ structure, allowed LaTeX section levels) but the core 7-step generation process is described as prose ('AI 分析章节主题...', '用硬编码模板包装成合法 LaTeX') with no executable code or commands; matches the 3 anchor of partial/abstract guidance. | 3 / 5 |
Workflow Clarity | A numbered 7-step sequence is present and a '校验/失败与恢复' section exists, but validation is described generically ('执行格式与结构验证', '完成后执行 Skill 已有的静态检查') without explicit pass criteria or a validate→fix→retry loop with concrete commands; per the rubric's destructive/batch cap and missing-checkpoint guidance, this caps at 3 even though apply-mode writes back to the project. | 3 / 5 |
Progressive Disclosure | Structure is reasonable: SKILL.md is an overview with a single references/README.md and a scripts/ bundle that is real and appropriately offloaded; references are one level deep and signaled (references/README.md, config.yaml, scripts/skill_controller.py named in the README). Minor gaps — the body does not inline-link the script files and leans on config.yaml without a clear pointer — keep it just below 5. | 4 / 5 |
Total | 13 / 20 Passed |