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.
A well-engineered instruction-only skill body: global constraints, confirmation gating, and failure-recovery rules are kept inline while detail is delegated to a real, one-level-deep reference set, and risky workflows all carry explicit dry-run/confirm checkpoints and recovery paths. The gaps are modest — a duplicated routing table vs the index file, a twice-stated default, and command templates that never show one complete invocation.
Suggestions
Conciseness: drop the SKILL.md 路由 table's overlap with dingtalk-tag-index.md (or reduce it to intents the index doesn't cover) and state the mention_only default only once, keeping the publish bullet focused on the save-draft remediation path.
Actionability: include one complete copy-paste invocation (e.g., a full `dws dingtalk-tag manage create` with --name/--description/--type and a `connect --channel dsh` example) so the common cases can be executed without opening a reference file.
Actionability: expand a few '...' routing-table entries (manage login, run) with their key flags so the routing layer alone disambiguates command choice.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The 47-line body is dense operational constraint with zero padding or explanations of known concepts, but has minor redundancy: the 路由 table overlaps the index file's 意图路由 table, and '创建时默认响应模式为 mention_only' appears in both the response-mode bullet and the publish bullet. Efficient with minor instances that could be trimmed fits anchor 4 rather than the every-token-earns-its-place bar of anchor 5. | 4 / 5 |
Actionability | Concrete commands with exact flags are given ('dws dingtalk-tag connect --agent-uuid ... --channel dsh', '--dry-run --format json', '--type open_code|local_agent', '--body-stdin --wait-delivery --employee-context'), but most commands are templates with '...' ellipses and no single complete copy-paste example with full arguments — those live in the reference files. Mostly executable with minor gaps fits anchor 4 rather than the copy-paste-ready coverage of anchor 5. | 4 / 5 |
Workflow Clarity | Explicit validation checkpoints ('先用 --dry-run --format json 展示计划', '确认后才加 --yes') and error-recovery feedback loops are present for fragile operations: create-partial-failure ('必须返回已创建的 agentUuid 和恢复命令；重试禁止再次执行 create'), connect failure ('检查 connect status 再恢复'), and rebind (unbind 成功后 connect, with retry rules for unknown/timeout/unbinding states). Destructive delete is marked 不可逆 with confirmation required, so the missing-validation cap does not apply; this matches anchor 5's explicit-validation-plus-feedback-loop pattern. | 5 / 5 |
Progressive Disclosure | The body routes through a verified one-level-deep reference set — all five cited files (dingtalk-tag-index.md, manage.md, manage-and-connect.md, capability.md, run.md) exist in references/ — each linked with a stated purpose ('生命周期读取 manage.md，自然语言创建/恢复/DSH 接入再读 manage-and-connect.md'). The body retains only routing, global hard constraints, and security invariants, which is the appropriate split. Clear overview with well-signaled one-level references matches anchor 5. | 5 / 5 |
Total | 18 / 20 Passed |