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 well-engineered, highly actionable skill body with verified executable commands, a clear mode-aware workflow, and exemplary progressive disclosure (exhaustive one-level index, read-only-what-you-need guidance). The main weakness is redundancy: mode behavior is explained in three separate sections, which inflates token cost and could be consolidated into the decision table plus workflow alone.
Suggestions
Consolidate the mode-branching logic that currently appears in 运行模式, 快速用法, and 提示词工作流 into one authoritative section, keeping the decision table and workflow steps as the single source of truth.
Remove the duplicated check-mode.js invocation examples (appears in both 运行模式 and 快速用法 0) — one occurrence plus the workflow step reference is sufficient.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly operational (mode table, env vars, naming rules, command examples), but mode-handling logic is repeated three times across "运行模式", "快速用法 0/5", and "提示词工作流" — e.g. '不要调用 generate.js' appears in Mode B while the decision table and workflow steps 7-B/7-C restate the same branching. This duplication could be consolidated, matching anchor 3 ('mostly efficient but could be tightened') rather than anchor 4. | 3 / 5 |
Actionability | Fully executable, copy-paste-ready commands that were verified against the actual script flags (`--prompt`, `--size`, `--quality`, `--promptfile`, `--image`, `--mask` all exist in generate.js/edit.js), plus concrete naming examples ('garden-gpt-image-2/prompt/live-commerce-ui-20260424-153045.md') and a precise mode decision table. This matches anchor 5: specific examples cover the common cases for all three modes. | 5 / 5 |
Workflow Clarity | The 8-step "提示词工作流" has a clear sequence with mode branching at step 7, a mandatory environment validation first step (`check-mode.js`), and explicit uncertainty handling ('Mode A 调脚本失败（401 / 网络 / 配额）→ 报错并询问'). It falls short of anchor 5 because error recovery is a single line without a retry/re-validate loop, and restating the workflow in three places risks mild ambiguity. | 4 / 5 |
Progressive Disclosure | SKILL.md is a pure overview plus a well-organized index: 18 category directories each listing its template files with one-line descriptions, all verified to exist in references/. References are one level deep (verified: no template references another file), navigation guidance is explicit ('只读取对应的具体模板文件，不要一次读整个 references/'), and detail is appropriately deferred to scripts/ and references/. This matches anchor 5. | 5 / 5 |
Total | 17 / 20 Passed |