Content
81%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 strong operational skill: fully executable commands, an explicitly gated decision tree with hard validation loops, and a real, complete reference bundle. Its weakness is token economy — script behavior and the faithful-clone default are repeated across two or three sections — plus a dangling flagship-case pointer.
Suggestions
Cut the duplicate script descriptions: keep the one-line purpose list in '内置脚本' and drop the behavior re-explanations already covered at each workflow step (e.g. asset-harvest's fonts.css output is stated in Step 2, the hard-rules section, and the script list).
Move the '逆向特效三件套纪律' (evidence grading, no-compensation, baseline-first) and the '资产与颜色保真' detailed rules into references/effect-extraction.md and references/assessment.md, leaving one-line rules plus pointers in SKILL.md.
Fix or remove the '旗舰案例 ./marbles-clone/' section — the directory is not part of the bundle — or restate it as an expected project output rather than a shipped example.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and directive but noticeably redundant: every script is described twice — once at its workflow step (e.g. Step 2's asset-harvest invocation with comments) and again in the '内置脚本' section ("scripts/asset-harvest.mjs：真浏览器网络栈全程滚动捕获并下载…"), and the asset-harvest behavior is explained a third time in '资产与颜色保真'. The 模式选择纪律 paragraph in Step 2.5 also re-argues the same faithful-clone default several ways. This is 'mostly efficient but includes some unnecessary explanation or could be tightened', not 4, because the duplication spans whole sections rather than isolated trimmable lines. | 3 / 5 |
Actionability | Guidance is copy-paste ready throughout: concrete bash blocks with full flags ("node scripts/audit-clone.mjs --project . --brand \"<原站品牌名>\" --recon RECON/original-recon.json --strict --out CLONE_AUDIT.md"), a decision table mapping each recon finding to a path, and exact expected values ("footer 是 rgb(17,17,17) 就写 #111111"). It also anticipates error recovery ("某张图下载失败就换 --recon 兜底源或从 RECON/network 捕获里捞"). Nothing is left as abstract direction. | 5 / 5 |
Workflow Clarity | The decision tree is strictly sequenced (Step 0 → 6 with '按顺序走，不许跳') and validation is explicit with feedback loops: "--strict…有硬伤 exit 2 —— 修完重跑，不通过不许交付", "浏览器真验证（硬要求，不许只看代码就说'应该能跑'）…诚实记录验证不了的部分", and the baseline-first gate ("最小原样可复现 RAW REPLAY → 逐帧比对通过 → 才允许重构"). These are exactly the validate→fix→retry checkpoints the rubric's anchor-5 example rewards. | 5 / 5 |
Progressive Disclosure | Eight well-signaled, one-level-deep references ("见 references/assessment.md", "详见 references/static-mirror.md"), all verified present on disk, plus 13 scripts each documented with a purpose. Not 5 because some material that belongs in references is inlined in the SKILL.md body — the ~40-line 逆向特效三件套纪律 and the 资产与颜色保真 rules would fit effect-extraction.md/assessment.md — and the '旗舰案例 ./marbles-clone/' is referenced but does not exist in the bundle, a dangling pointer. | 4 / 5 |
Total | 17 / 20 Passed |