Content
78%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 extremely lean, gated, mostly executable workflow overview with real validation checkpoints — strong on token efficiency and action. The main weakness is navigation against the actual bundle: the majority of referenced files do not resolve here (bare filenames, no paths), and the reference files that do exist are never surfaced from SKILL.md.
Suggestions
Give every file in the step chains an explicit path (e.g. prompts/claim_strategy.md, tools/check_design_views.py) so each link resolves unambiguously.
Link the existing bundle references from SKILL.md — e.g. point the specification step to references/disclosure_to_spec_map.md and mention references/schemas/ and references/promo_terms.yaml in 产出物 or 护栏 — so present files are discoverable.
Show the invocation for tools/emit_application_docx.py with its arguments, matching the copy-paste style already used for material_gate.py.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | All ~30 lines are operational — no concept explanations, no filler; e.g. "仅缺文件时终止" states a stop rule in six characters and the 何时用 section defines the trigger contract in two sentences. Every token earns its place; nothing explains what Claude already knows. | 5 / 5 |
Actionability | Steps are mostly executable and concrete: "跑 tools/material_gate.py --case-dir <交底目录>;退出码 2 则停", "跑本包 check_design_views.py", "本包 tools/emit_application_docx.py 出 Word", plus the repo-wide path "python skills/patent-application/tools/…". Not a 5: emit_application_docx.py is invoked without arguments, and the "claim_strategy.md → claims_builder.md → …" chains rely on implied Read semantics rather than an explicit verb. | 4 / 5 |
Workflow Clarity | A clear numbered 1–6 sequence with explicit validation checkpoints (material gate with exit-code-2 stop, the design-view check, the consistency.md step), a bypass for the revision case, and a separate 问题清单 checklist. Not a 5: recovery after a failed check is only "停" — no fix-and-retry feedback loop. The destructive/batch cap at 3 does not apply since outputs always go to a new timestamped directory. | 4 / 5 |
Progressive Disclosure | The body is a well-structured one-level-deep overview, but scored against the actual bundle: most referenced paths (prompts/guardrails.md, tools/material_gate.py, check_design_views.py, emit_application_docx.py, issues.md, and the whole claim_strategy.md → … → consistency.md chain) are not present in this bundle, and those step files are cited as bare filenames with no directory. Meanwhile existing bundle references — references/disclosure_to_spec_map.md, references/promo_terms.yaml, references/schemas/ — are never linked from SKILL.md, so present content is undiscoverable from the overview. | 3 / 5 |
Total | 16 / 20 Passed |