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.
A strong, dense instruction body: it routes every user intent to a concrete command, states exact flags and validation rules, includes dry-run/confirm checkpoints and failure-recovery guidance for destructive operations, and discloses progressively into five real, purpose-labeled reference files. Its weaknesses are repetition between the intro, routing table, and constraint bullets, overloaded multi-clause sentences, and the absence of one complete example command for the most common create flow.
Suggestions
Deduplicate guidance that appears in multiple places — the mention_only default (intro paragraph plus two constraint bullets) and the unbind/connect flows (routing table plus four constraint bullets) — keeping policy in one authoritative location.
Add one complete, copy-paste-ready example of the most common operation, e.g. `dws dingtalk-tag manage create --name ... --description ... --type qwenwork --dry-run --format json`, so the required-argument contract is visible without opening references.
Break the longest multi-clause constraint bullets (e.g. the --type / --snapshot bullet and the 换绑 bullet) into single-rule lines or a short numbered sequence for the create → publish → connect path, making the order and checkpoints easier to follow.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is information-dense with no explanation of concepts Claude already knows — every bullet carries product-specific rules (e.g. "`digitalTagEmployeeProfile.type` ... 仅支持 `qwenwork`、`local_agent`"). However, there is noticeable repetition: the `mention_only` default appears in two bullets and the intro paragraph, and unbind/connect flows appear in both the routing table and the constraint bullets, so it sits at anchor 4 ('minor instances that could be trimmed') rather than 5. | 4 / 5 |
Actionability | Commands are concrete with exact flags and values (`--dry-run --format json`, `--agent-uuid ... --channel dsh`, `--daemon --alwayson`, `configString.mcpServers.<名称>.type`), which is mostly executable guidance. It falls short of anchor 5 because commands use `...` elision and the most common operation (create with its required name/description/type arguments) never appears as a complete example — full syntax is delegated to references, leaving minor gaps. | 4 / 5 |
Workflow Clarity | Multi-step flows have explicit checkpoints and feedback loops: "先用 `--dry-run --format json` 展示计划" then "确认后才加 `--yes`", "旧机器 `connect unbind` 成功后,新机器 ... connect", and failure recovery like "connect 失败后 ... 检查 `connect status` 再恢复" and "重试禁止再次执行 create". Destructive operations do have validation, so no cap applies, but sequences are scattered across bullets rather than consolidated, leaving minor gaps versus anchor 5's explicit checklists. | 4 / 5 |
Progressive Disclosure | The body is a concise overview that clearly signals one-level-deep references, all of which exist in the bundle: "先用 [dingtalk-tag-index.md] 路由;生命周期读取 [manage.md] ... 能力资源读取 [capability.md];执行排障读取 [run.md]" (plus manage-and-connect.md). Content is appropriately split (a ~45-line body routing to five purpose-labeled reference files), matching anchor 5's clear overview with well-signaled references; it is not anchor 4 because navigation is complete and unambiguous. | 5 / 5 |
Total | 17 / 20 Passed |