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.
An exemplary router/index skill body: when-conditioned reference map, explicit load-discipline rules, a numbered precedence order with concrete per-overlap resolutions, and nothing inlined that belongs in reference files. The workflow and validation steps are clear but stop short of concrete commands and an error-recovery loop, and a few generic workflow lines could be trimmed for token efficiency.
Suggestions
Trim generic workflow lines that restate default behavior (e.g., "Inspect the relevant implementation and trace the affected interaction", "Make the smallest complete change that satisfies the request") and keep only policy specific to these trees, such as the state checklist and polish-review gating.
Name the actual check commands per tree in the Validate step (e.g., the lint/type/test scripts from each tree's package.json) instead of the generic "Run the smallest relevant lint, type, and test checks".
Add a fix-and-re-run feedback loop to the Validate step and specify how to verify rendered behavior (e.g., the dev-server command for each tree), which would also raise workflow clarity.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and almost entirely project-specific (tree paths, precedence order, per-file ownership) with no explanations of concepts Claude already knows, but a few workflow lines restate default behavior ("Make the smallest complete change that satisfies the request", "Inspect the relevant implementation") and could be trimmed. Fits anchor 4 (efficient with minor trimmable instances) rather than 5, since not quite every token earns its place. | 4 / 5 |
Actionability | The "Read this | When" reference map, numbered precedence order, and per-overlap resolutions ("Apply the shadcn rule to component wiring, the polish rule to visual detail") are concrete decision rules, and deferring to each tree's package.json is explicitly justified. Falls short of anchor 5 because the Validate step names no actual commands ("Run the smallest relevant lint, type, and test checks") and "verify the rendered behavior" specifies no method. | 4 / 5 |
Workflow Clarity | Five clearly sequenced steps (Understand → Select → Implement → Validate → Report) with an explicit Validate step and a concrete state checklist ("loading, empty, error, disabled, responsive, keyboard, focus, and reduced-motion states"). Not anchor 5 because there is no fix-and-re-run feedback loop and the checks are unnamed; not 3 because checkpoints are present rather than implicit. | 4 / 5 |
Progressive Disclosure | The body is a pure index: a when-conditioned reference table, explicit anti-loading rules ("do not load it by default", "Do not load motion references for work with no motion concern"), and inline enumeration of each area's sub-files so navigation stays effectively one level deep with no blind pointer-chaining. Note: no bundle files were present in the evaluation environment to verify the referenced paths exist, so this scores the reference structure as written. | 5 / 5 |
Total | 17 / 20 Passed |