Content
85%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.
An effective, disciplined router skill: crisp gated routing workflow, concrete tool identifiers and commands, delivery/evidence checklists, and an explicit anti-pattern table. Its two real weaknesses are dangling refs/*.md references (none of the four referenced files exist in the bundle) and noticeable triplication of the routing mapping across three sections.
Suggestions
Ship the four referenced files (refs/playwright-mcp.md, refs/playwriter.md, refs/agent-browser.md, refs/pinchtab.md) in the bundle — or point the refs table at the actual existing paths — so the skill's one-level-deep navigation resolves.
Merge the overlapping '路由矩阵' and '读取哪些 refs' tables into a single table (场景 | 后端 | MCP ID | 状态 | ref), since both encode the same mapping and one can carry the ref column.
Remove the verbatim repetition of questions 1–2 ('真的需要浏览器吗', '目标是 localhost 吗') between '执行前四问' and '默认路由顺序' by having the routing order reference the four questions instead of restating them.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and assumes Claude's competence — no explanations of what browsers or MCP are — and states '而不是在这里重复厂商文档' / '厂商细节压到 refs/'. It falls short of 5 because routing information is presented three times (默认路由顺序 prose, 路由矩阵 table, 读取哪些 refs table) and the first two routing questions are repeated verbatim in both '执行前四问' and '默认路由顺序'. | 4 / 5 |
Actionability | For an instruction-only router skill, guidance is fully concrete: specific MCP IDs ('playwright', 'pinchtab'), tool prefixes ('mcp__claude-in-chrome__*'), install commands ('npm i -g agent-browser', 'npx @playwright/mcp@latest'), exact ref paths, and a precise failure fix ('用 pinchtab_eval + window.location.href 导航'). Per the rubric's code-vs-instruction note, absence of code is not penalized when guidance is this actionable. | 5 / 5 |
Workflow Clarity | The '默认路由顺序' is a gated 7-step decision sequence with explicit conditions; '交付要求' is a delivery checklist including evidence and login-session accountability; 'Common Mistakes' provides error→consequence→fix recovery loops; and step 6 includes an explicit validation ('使用前先做本机验活'). This matches anchor 5's sequence + checkpoints + checklist + feedback loops; it is above anchor 4 because recovery and validation are explicit, not merely implied. | 5 / 5 |
Progressive Disclosure | The structure is designed correctly — a lean overview with '读取哪些 refs' pointing one level deep to refs/playwright-mcp.md, refs/playwriter.md, refs/agent-browser.md, refs/pinchtab.md — but none of these files exist in the bundle (no refs/, references/, scripts/, or assets/ directories are present), so every deep link is dangling and navigation breaks in practice. It is above anchor 2 (structure and signaling are good, nothing monolithic) but below anchor 4 because referenced paths do not resolve to real files. | 3 / 5 |
Total | 17 / 20 Passed |