Content
58%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.
The body delivers an exceptionally actionable, well-sequenced workflow with genuine validation gates and error-recovery loops — among the strongest operational guidance a skill can have. But it pays for that with severe token bloat: ~27k tokens of dense, repetitive, exception-driven prose loaded on every invocation, much of it reference material that belongs in on-demand files. Actionability and workflow clarity are top-tier; conciseness and progressive disclosure pull the overall quality well down.
Suggestions
Move the Griffel shorthand list, longhand expansion rules, and iterator-compatibility rules into the referenced rules file, keeping only the check command and 'fix what it reports, rerun until exit 0' inline — this alone removes ~100 lines that duplicate the checker's own output.
Consolidate the table-publish scoping logic, which is stated at length twice (Phase 2 '--publish' section and Notes & limits), into one authoritative location and cross-reference it.
Move the per-feature 'minimumPluginVersion' gates and the declared form/dashboard/rule/BPF adoption and halt minutiae into the advanced schema reference, leaving one-line decision rules ('only when the maker asks; explain the approval') in the toolbox bullets.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | A ~1281-line / ~27k-token body that is severely padded: the Griffel section enumerates ten shorthands and their longhand expansions inline (~40 lines) instead of trusting the checker's own output; the table-publish scoping logic is stated at full length twice (Phase 2 and Notes & limits); single sentences run 300+ words (the declared-form bullet); and "minimumPluginVersion: \"2.21.0\"" is repeated inline ~15 times. Not 2: the verbosity is systemic and repetitive rather than a few padded sections — 'severely verbose ... heavily padded' is the closest anchor, and the guideline 'penalize verbosity even if accurate' applies fully. | 1 / 5 |
Actionability | Every step ships a copy-paste-ready command with flags, expected output, and failure semantics: `build-model-app.js --env <envUrl> --spec @<working-dir>/app-spec.json --apply --verify`, "exits 3 and leaves app-spec.json untouched", `{ ok, planPath, pages: [{ name, key, file, dataMode, intent }] }`, and a complete verbatim worker-dispatch contract. Not 4: commands are fully executable, common cases are covered, and even exit codes and stderr behavior are specified. | 5 / 5 |
Workflow Clarity | The workflow is clearly phased (Phase 0 → 1 → 1.5 → 2 → 3) with explicit validation at every gate: the lint hard gate ("errors block"), single plan-mode approval on the dry-run plan, the atomic promote-intent-pages check, `--verify` after apply, and teardown dry-run by default with `--allow-destructive` required. Feedback loops are spelled out ("fix what it reports and run it again", halt codes each with a recovery hint, "run the full build again" for recovery). Not 4: checkpoints and error-recovery loops are explicit throughout, including for the destructive teardown and batch data-fix paths. | 5 / 5 |
Progressive Disclosure | References are well signaled and one level deep ("[references/authoring-flow.md]", "read app-spec-schema-advanced.md once your design calls for a conditional feature"), but hundreds of lines of reference-grade detail are inlined in SKILL.md itself — Griffel/iterator rules, publish scoping, declared form/dashboard/rule/BPF adoption minutiae, teardown ordering — and none of the referenced files (references/, scripts/, samples/) exist in the reviewed bundle. Not 3: the inlining is not one or two sections but the dominant structure of a 1281-line file that functions as its own reference manual; not 1: sections are well organized and references that are made are clearly signaled. | 2 / 5 |
Total | 13 / 20 Passed |