Content
92%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 unusually disciplined instruction-only skill body: no filler, no explanations of known concepts, exact command syntax, and validation/feedback loops woven through every risky path (formula verification, read-back asserts, high-risk consent gates, async polling). Structure is a genuine progressive-disclosure hub — all 20 references exist and are routed per action, with scripts correctly surfaced as helpers. The only realiable criticism is density/length: several rules and the flag-convention section pack a lot into single passages that could be tightened or split out.
Suggestions
Tighten 准则 2 and 准则 4, which restate style-inheritance and read-back details already delegated to references/lark-sheets-write-cells.md — keeping only the principle in the body and pointing to the reference would cut the longest passages.
Consider moving the cross-platform shell-adaptation table (PowerShell/cmd.exe payload forms) into a small reference or the shared sibling skill, since it is only needed on non-POSIX platforms and is dead weight in the common bash path.
The AI-formula blockquote under the editing principles duplicates trigger/verification rules that also live in references/lark-sheets-formula-translation.md and formula-verify; a short rule plus a pointer would preserve the routing at lower token cost.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body contains essentially no padding or explanation of concepts Claude already knows — every line is lark-cli-specific rules, flags, and caveats — but it is long (~28KB) and a few sections strain clarity: the 200+ character single-sentence rules (准则 2 and 4), the AI-formula blockquote, and the flag-convention/shell-adaptation material that could partially live in a reference. This sits between anchor 3 (some tightening possible) and anchor 5 (every token earns its place): efficient overall with trimmable over-length, matching anchor 4 rather than 5 because the editing-principles section reiterates details (style re-read, insert-vs-extend style inheritance) that also appear in the scenario table and are delegated to references. | 4 / 5 |
Actionability | Guidance is fully executable: exact commands with flags ("lark-cli sheets +csv-get --url \"https://.../sheets/shtXXX\" --sheet-name \"<真实表名>\" --range \"A1:F30\""), exact syntax rules ("显式值只用 --flag=true 或 --flag=false"), concrete shell-per-platform payload forms, and per-action command selection in the scenario table. Anchor 5 (copy-paste ready commands covering the common cases) fits; anchor 4 would imply minor gaps in executable detail, but even edge cases like bash history expansion ('\'' escaping) and cmd.exe quoting are covered concretely. | 5 / 5 |
Workflow Clarity | A clear sequence is stated ("流程：了解结构 →…→ 读数据 → 原子工具写入 → 按用户点名项回读验证 → 在线交付") and validation checkpoints are pervasive: read-back asserts on first/middle/last cells and named items ("返回 ok 只表示请求成功"), formula closure via "+formula-verify --exit-on-error" with sentinel records, retry-then-downgrade loop ("试错 3 次仍失败可降级静态值…交付说明写明"), async status polling for AI formulas and history revert, and a dry-run → user consent → --yes gate for high-risk batch/destructive commands. This matches anchor 5 (explicit validation, feedback loops, effective checklists); the destructive/batch cap at 3 does not apply because validation is the body's strongest feature. | 5 / 5 |
Progressive Disclosure | The body is a routing overview: a scenario→command table whose '动手前读' column names the exact reference to read per action, and a two-group References catalog covering all 20 files — every referenced path was verified to exist in the bundle. References are one level deep (cross-pointers between references are shallow navigation aids), scripts are surfaced from the relevant references, and the three unreferenced scripts are declared helper modules ("Range and coordinate helpers for Lark Sheet scripts", "Pure sizing heuristics shared by Lark chart helper scripts"), so the bundle structure is coherent. This matches anchor 5 (clear overview, well-signaled one-level-deep references, easy navigation) rather than anchor 4, which would require organization gaps or unclearly signaled references. | 5 / 5 |
Total | 19 / 20 Passed |