Content
77%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.
This is a highly actionable, operationally precise skill: executable Code Mode and runtime snippets, exact limits, explicit validation loops, and well-gated destructive operations. Its weaknesses are redundancy — several rules (system fields, memory-file maintenance, tableId warnings) are each repeated 3–5 times, with inconsistent bilingual remnants — and the absence of any progressive disclosure, with the full API reference inlined in a 565-line SKILL.md.
Suggestions
Consolidate the repeated rules into single canonical statements: the 'never write system fields' rule appears in five places, the MICRO-APP.md/.magicbase/migrations.json maintenance note in four, and the tableId-vs-table_key warning in two — state each once in its most relevant section and reference it elsewhere.
Split the detailed material into one-level-deep reference files (e.g., references/schema-tools.md for the schema tool reference, references/runtime-api.md for the window.Magic.db API and datetime formats, references/permissions.md for the permission matrices), leaving SKILL.md as a lean overview with clearly signaled links.
Translate or remove the leftover Chinese fragments ("表 ID", "过滤条件", "排序规则", "返回字段列表", "页码", "每页行数", "分页结果...不要使用 rows", "超时:30 秒", "获取单行", "获取relation list", "可能的错误") so the skill uses one consistent language and saves tokens.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The content is dense with genuinely novel domain rules (no basic-concept teaching), but the same rules are restated many times: "Do not write system fields such as `created_by`, `organization_code`, `created_at`" / "must not put system fields into the `data` object" / "never write system fields" appears in at least five places, the `MICRO-APP.md` + `.magicbase/migrations.json` maintenance note appears four times, and the tableId-vs-table_key warning and the `select`-must-include-`created_by` rule are each duplicated. Leftover Chinese fragments ("表 ID", "过滤条件", "分页结果", "超时:30 秒", "获取单行") add inconsistent token overhead. This fits the 3 anchor — mostly efficient but could be tightened — rather than the 2 anchor, since the padding is duplication of novel rules rather than explanation of things Claude already knows. | 3 / 5 |
Actionability | Guidance is fully executable and copy-paste ready: complete Code Mode snippets ("from sdk.tool import tool; result = tool.call(\"query_magicbase_tables\", {}); if not result.ok: raise SystemExit(result.content)"), full runtime signatures with concrete calls ("window.Magic.db.createRow(TABLE_ID_FROM_MAGICBASE_TOOL, {...}, [\"id\", \"name\", ...])"), a ready admin guard and `canEdit` implementation, exact operational limits ("at most 100 numeric row IDs", "at most 200 rows per call", "30 秒(其他操作为 15 秒)"), and real error messages in the error-handling example. Specific examples cover the common cases (create table, query/update, batch get, permission JSON), matching the 5 anchor. | 5 / 5 |
Workflow Clarity | The schema-first workflow is explicitly numbered with a hard checkpoint ("1. Call `query_magicbase_tables`... 4. Generate HTML only after you have a real `table.id`"), and every snippet carries a validation/feedback loop ("Always check `result.ok` before using the result"). Destructive and batch operations — the cap trigger in the rubric — are properly gated: "destructive; only after explicit user confirmation", "Before deleting, call `query_magicbase_rows` to resolve the exact real row IDs and show the user the deletion scope", plus retry guidance ("If the tool rejects `dynamic_permissions` with 'expected object', retry with the same permission intent as an object") and a pending-migration repair loop. Explicit validation, error-recovery loops, and decision checklists (the four permission questions, the enforceable/ui_only/requires_backend classification) match the 5 anchor. | 5 / 5 |
Progressive Disclosure | There are no bundle files at all — the entire ~565-line skill lives in SKILL.md, including full "Schema tool reference" and "MagicBase Runtime Database API" sections, a datetime-format catalog, and permission matrices that clearly belong in one-level-deep reference files. Section headers are clear and navigation within the single file is feasible, which places it at the 3 anchor ("some structure... content that should be separate is inline") rather than the 2 anchor's minimal-structure example; it cannot reach 4–5 because virtually all detail is inlined with no references to offload it. | 3 / 5 |
Total | 16 / 20 Passed |