Content
70%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 highly actionable, well-sequenced workflow with strong validation and feedback loops, but the single-file structure inlines large lookup tables that belong in reference files, and the filter/description rules are restated across multiple sections. Consolidating duplicated rules and moving the emoji/component/compliance tables to references/ would tighten the skill significantly.
Suggestions
Consolidate the filter rules that appear near-verbatim in Phase 1 ('过滤规则:哪些 PR 不写入 changelog') and Phase 2 ('过滤规则') into a single authoritative section, and drop the restatement of description rules in '注意事项'.
Move the emoji spec table, commit-type→verb mapping, 55-entry component name list, and compliance rule table into a references/ file (e.g. references/changelog-rules.md), keeping SKILL.md as a lean workflow overview with a clearly signaled link.
Fix the subagent batch example so it reflects how gh actually accepts PR numbers (e.g. a loop or gh pr view per number inside each subagent) so the parallel-collection step is copy-paste executable.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly dense project-specific rules (emoji table, commit-type→verb mapping, compliance table) rather than explanations of things Claude already knows, but there is real duplication: the filter rules appear nearly verbatim twice ('以下改动不应出现在 changelog 中' in Phase 1 and '过滤规则' in Phase 2), the description/signature rules are restated in '注意事项', and the 22-step '完整交互流程' recaps every phase already described. Not a 2 because nearly all content is non-obvious, antd-specific policy; not a 4 because roughly a fifth of the file is restated material that could be trimmed. | 3 / 5 |
Actionability | Concrete executable commands throughout: 'git tag --list | grep -v -E "(experimental|alpha|resource)" | sort -V | tail -20', 'gh pr view <pr-number> --json title,body,author,number', 'npm run lint:changelog', plus exact output formats for ~changelog.md entries. Not a 5 because the subagent diagram shows 'gh pr view 56001 56002 ...' which gh pr view does not accept as multiple arguments, and 'cat >> ~changelog.md' is offered as an alternative without a worked example. | 4 / 5 |
Workflow Clarity | The 3-phase workflow is explicitly sequenced, summarized again in a numbered 22-step flow, and includes validation checkpoints with a feedback loop for the batch write operation: preview confirmation before writing ('[预览完整内容] [直接写入] [取消]'), and '运行 npm run lint:changelog 校验,有报错则修正并重新写入,直到 lint 通过' — an explicit validate→fix→retry loop for the destructive file writes. | 5 / 5 |
Progressive Disclosure | No bundle files exist (references/, scripts/, assets/ are absent), so all ~530 lines live in SKILL.md. In-file structure is well sectioned, and there is one clearly signaled external reference ('参考项目根目录 AGENTS.md 中的 Changelog 规范'), but large reference material — the 55-entry component name list, the emoji spec table, and the compliance rule table — is inlined where a references/ file would keep the overview lean. Not a 2 because the sections are well organized and the single reference that does exist is properly signaled. | 3 / 5 |
Total | 15 / 20 Passed |