Content
63%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.
This is a sophisticated, well-structured workflow skill with explicit phase completion criteria, feedback loops, and a thoughtfully designed on-demand reading system. Its main weaknesses are density — long run-on paragraphs and rules repeated across sections that inflate the token budget — and the fact that all referenced reference files are missing from the bundle while their content is partly duplicated inline.
Suggestions
Move the detailed protocol variants (SDK Run 路径 continuation modes, Compat/legacy handling in '后续指令') into the referenced files (reference/commands.md, reference/recovery.md) and keep only the decision rules inline — the body currently duplicates what the references should carry.
Break the multi-clause run-on paragraphs (rules section, Builder step) into short bullet points, and deduplicate rules stated in both '必须遵守的规则' and the phase sections (e.g., failure-not-passing, artifact-position checks).
Ensure the referenced bundle files (reference/*.md and ../comet/reference/*.md) actually ship with the skill, or inline the minimal needed content — every '按需读取' link currently points to a file absent from the bundle.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body avoids explaining concepts Claude already knows, but it is densely packed with run-on paragraphs (e.g., the multi-clause rule paragraphs in '必须遵守的规则' and the Builder step in 'SDK Run 路径') and repeats rules across sections (e.g., failure-not-passing appears in both the rules and Verify; '不能用另一文件中的同类内容代替' appears in both Build and Verify). It could be tightened noticeably without losing information, matching anchor 3 rather than 4. | 3 / 5 |
Actionability | Concrete, executable guidance is present throughout: specific commands ('comet native select <change-name> --json', 'comet native status --json', 'native new --task'), flags ('--review current-session', '--accept-on-pass <用户授权摘要>'), files ('comet-state.yaml', '.comet/config.yaml' hook.allow_paths), and response types. Minor gaps remain: several actions defer exact syntax to Runtime responses or '--help', and some policy sentences ('由当前 Agent 根据不确定性和用户要求决定') are abstract, keeping it below anchor 5. | 4 / 5 |
Workflow Clarity | The multi-phase workflow (开始或恢复 → Shape → Build ↔ Verify → Archive → 后续指令) is clearly sequenced with explicit validation checkpoints via per-phase '完成标准' sections and feedback loops ('未通过则回 Build 修复并重新提交', failure/timeout never counts as pass, recovery references for interruptions). It falls short of anchor 5 because interleaving three runtime variants (SDK, Compat, legacy) inside single paragraphs makes some sequences hard to follow on first read. | 4 / 5 |
Progressive Disclosure | The '按需读取' section is genuinely well-designed — each reference link carries an explicit read condition (e.g., '状态包含 childSummary 时...读取 Supervisor 协作'), one level deep. However, none of the referenced files (reference/commands.md, reference/recovery.md, ../comet/reference/*.md, etc.) exist in the bundle, so the disclosure targets are unverifiable, and ~135 lines of dense protocol detail that belongs in those references (continuation modes, SDK/Compat protocol variants) are inlined in SKILL.md, matching anchor 3. | 3 / 5 |
Total | 14 / 20 Passed |