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 strong router-style SKILL.md: commands are copy-paste executable, workflows are explicitly sequenced with thorough validation and confirmation gates for destructive operations, and nearly all domain detail is properly deferred to a well-organized references/ directory. The main weaknesses are a missing core bundle file for the creative-design path (a dead link for the entire html app_type), an inlined member-management section inconsistent with the file-per-domain pattern, and duplicated +delete confirmation rules that add token weight.
Suggestions
Add the missing creative-design/creative-design.md to the bundle (or repoint the four references in the 意图路由 table, the app_type 判定表, and the html 主链路 bullet to its actual location) — the app_type=html routing currently dead-ends.
Move the「应用协作者与协作权限设置」section's detail into a references/lark-apps-member.md file (matching the cache/db/env/role pattern), keeping only the routing row and the top risk rules in SKILL.md.
In the access-scope routing row, replace the vague "对应 access-scope reference" with explicit links to references/lark-apps-access-scope-get.md and references/lark-apps-access-scope-set.md, and deduplicate the +delete confirmation protocol between the routing table and「禁止预授权判定底线」②.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is a dense router that assumes Claude's competence and defers domain detail to references, with no explanation of concepts Claude already knows. Not 5: the +delete confirmation rules are fully duplicated (the routing-table row and "禁止预授权判定底线"② both spell out the same show-then-wait-for-confirmation protocol), the member-management section inlines reference-grade detail, and ② is a single sprawling multi-clause paragraph that could be split and tightened. | 4 / 5 |
Actionability | Fully executable, copy-paste-ready commands throughout: "lark-cli apps +member-add --app-id <app_id> --member-type openid --member-id ou_xxx --perm view --dry-run --as user", "lark-cli apps +get --app-id <meta_token> -q '.data.app.app_id'", "lark-cli apps +metric-list --app-id <app_id> --metric requests --since <range> --as user", with concrete error subtypes ("missing_scope", "feature_not_available", OpenAPI code 3340005) and exact typed-ID examples (ou_/oc_/od-). Common cases are covered with specific flag guidance (e.g., "不传 --series 会同时返回 total/error"). | 5 / 5 |
Workflow Clarity | Multi-step processes are explicitly sequenced with validation checkpoints: "进意图路由前先判这步" orders dev-path selection before routing; app_id resolution is an ordered 1-2-3 fallback; release flow specifies polling "+release-get 轮询 finished 给 online_url / failed 给 error_logs" with the PENDING/approval rule; git failures have a feedback loop ("先执行 lark-cli apps +git-credential-init…再重试原 git 命令…如果刷新凭证也失败,停止并向用户报告"). Destructive and batch operations (+delete, +env-delete, +cache-clear, DB ops, batch member removal) all carry explicit --dry-run/confirmation gates, so the destructive-ops validation cap does not apply. | 5 / 5 |
Progressive Disclosure | Structure is mostly excellent — 33 real one-level-deep reference files clearly signaled via the routing table's "按需读取" column — but the bundle check finds real defects: "creative-design/creative-design.md" is the designated load target for all four mentions of the html/creative path (one of three app_types) yet does not exist in the bundle, a dead end for a primary use case; the "应用协作者与协作权限设置" section inlines ~15 lines of reference-grade detail that every other domain (cache, db, env, role) keeps in its own references/ file; and one routing row says "对应 access-scope reference" without naming the actual files. Not 4 because a broken core reference is more than a minor organization gap; not 2 because the remaining 30+ intents route cleanly to well-organized, clearly signaled files. | 3 / 5 |
Total | 17 / 20 Passed |