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.
A well-designed router skill with genuinely concrete tool IDs, commands, pre-flight questions, and delivery requirements, and a correct keep-vendor-detail-in-refs architecture. Its two real weaknesses are section-level redundancy (the same routing decision restated six ways) and the fact that every referenced refs/*.md file is missing from the delivered bundle, which both weakens actionability and breaks progressive disclosure.
Suggestions
Collapse the redundant routing views: 默认路由顺序 already encodes the decisions in 先按目的选入口 (including the verbatim-repeated "真的需要浏览器吗" question), so merge 入口/路由矩阵/组合打法/读取哪些 refs/下一步 into the routing order or a single matrix.
Ship the four referenced refs/*.md files (playwright-mcp.md, playwriter.md, agent-browser.md, pinchtab.md) in the bundle, or inline the minimal per-backend usage notes, so routing decisions don't dead-end at missing files.
Add an explicit post-execution validation step to 交付要求, e.g. verify the evidence artifact (screenshot/text/log) exists and is non-empty before reporting completion.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly tight (tables, vendor detail pushed to refs), but the routing decision is restated across ~6 overlapping sections: "真的需要浏览器吗?…不要默认上浏览器" appears verbatim in both 执行前四问 and 默认路由顺序, and 路由矩阵/组合打法/refs 表/区别表/下一步 re-cover the same routing. This is more than minor trimming, so it sits at anchor 3 rather than 4, but there is no fluff or explanation of concepts Claude already knows, so not 2. | 3 / 5 |
Actionability | Concrete, executable routing guidance: "npm i -g agent-browser", "MCP ID: playwright", "工具前缀 mcp__claude-in-chrome__*", and the PinchTab fix "用 pinchtab_eval + window.location.href 导航", plus a four-item delivery checklist. Not 5 because execution detail is deferred to refs/*.md files that are absent from the bundle, so the paths dead-end; well above anchor 3 since what is present is specific, not pseudocode. | 4 / 5 |
Workflow Clarity | Clear sequenced workflow: 执行前四问 pre-flight checklist, numbered 默认路由顺序 decision sequence with explicit conditions, 交付要求 delivery checklist, and Common Mistakes failure→fix pairs. Anchor 4 rather than 5 because there is no explicit post-execution validation loop (e.g., confirming the promised screenshot/evidence artifact was actually captured before reporting). | 4 / 5 |
Progressive Disclosure | References are well-signaled and one level deep (refs table lists refs/playwright-mcp.md, refs/playwriter.md, refs/agent-browser.md, refs/pinchtab.md with scenarios), but the bundle contains no refs/ directory — none of the four referenced files exist, so the disclosure chain is broken. That is a larger gap than anchor 4's "minor organization gaps", landing at 3; not 2 because the SKILL.md itself is well structured and nothing that belongs in refs is inlined. | 3 / 5 |
Total | 14 / 20 Passed |