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.
The body is an efficient, highly actionable protocol document with a well-validated workflow: explicit pre-execution checks, structured failure outcomes, and resume semantics. Its one real defect is that the single external reference is a dangling path, so progressive disclosure cannot be relied on as written.
Suggestions
Ship the referenced examples/llm-replacement.md in the bundle, or remove the pointer if the example will not be provided — a dangling reference breaks navigation at the exact moment the reader needs it.
Trim the duplicated Decision enumeration (硬规则 4 vs 标准流程 7) to a single authoritative list.
Give one concrete invocation form for a `research_` tool (e.g., the actual command syntax for `research validate`) so the workflow's first executable step is unambiguous.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The ~90-line body is dense and lean: no concept explanations Claude already knows, no padding, and every section (触发门禁, 硬规则, 标准流程, minimal Spec, 输出契约) earns its tokens. The only trim candidate is the Decision enumeration appearing in both 硬规则 4 and 标准流程 7, which is too minor to drop below the 'lean and efficient' anchor. | 5 / 5 |
Actionability | Concrete tool calls with a stable `research_` prefix (research_validate, research_create, research_inspect, research_execute, research_continue, research_status, research_compare, research_evidence) in a numbered order, plus a near-complete copy-paste JSON Spec example and a literal output contract. Not a 5 because the Spec contains placeholders ("..." input, "path/to/adapter.py") and the body gives no invocation syntax for the `research_` tools, delegating that to the runtime. | 4 / 5 |
Workflow Clarity | The 标准流程 is a clear 7-step sequence with an explicit pre-execution validation checkpoint (step 3: 用 research_inspect 检查…缺关键条件时停止执行并返回结构化缺口), resume handling (research_continue instead of duplicating runs), no-hot-polling status checking, and a structured failure vocabulary (UNSUPPORTED/INVALID/INCONCLUSIVE) plus explicit multi-device conflict handling (硬规则 7). This matches the anchor's explicit validation steps and error-recovery feedback loops. | 5 / 5 |
Progressive Disclosure | Sections are well organized and content is appropriately inline for a ~90-line skill, but the body's only external reference — "相关示例见 examples/llm-replacement.md" — points to a file that does not exist in the bundle (no references/, scripts/, assets/, or examples/ directories). A clearly signaled yet dangling reference leaves navigation incomplete, which fits the 'structure present but organization/reference gaps' anchor better than the 'well-signaled one-level-deep references' anchor. | 3 / 5 |
Total | 17 / 20 Passed |