Content
88%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 is a lean, high-signal instruction set with explicit workflows, validation checkpoints, and error-recovery guidance for both report and fix modes. Its weaker spots are the deliberately unspecified `audit_live` invocation syntax and the named external flows/prompts that have no in-package navigation, which slightly limit actionability and progressive disclosure.
Suggestions
Provide a concrete example invocation (or discovery step) for the primary `audit_live` flow so the preferred path is as executable as the fallbacks.
Since no bundle files exist, either add a short pointer to where the `audit-live-page` prompt and flow definitions live, or trim the references to installed-tool discovery so navigation doesn't dead-end.
Consider moving the detailed report-format template into a bundled reference file (with a one-line pointer) to keep SKILL.md a tighter overview.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense, imperative, and assumes Claude's competence — e.g., "Sample representative files; don't open everything blindly", "Many low-impact violations of one rule are often a single root-cause fix", "Don't default-to-fix when the user only asked for an audit". No concept explanations (never explains WCAG, CDP, or libraries) and the report template is functional rather than padding, matching the 5 anchor. | 5 / 5 |
Actionability | Mostly executable guidance with concrete parameterized calls (`format: "compact"`, `audit_diff({ audit_name: "before" })`, `explain_rule({ id: "<rule-id>" })`), an exact report template, and an explicit fallback chain (grep stable hooks → visible text → tree position). Minor gaps: the primary `audit_live` invocation has no concrete syntax ("launch behavior depends on the installed integration") and the `audit-live-page` prompt's location is unspecified, so it lands on 4 rather than 5. | 4 / 5 |
Workflow Clarity | Both modes have clear sequences with explicit validation checkpoints and error recovery: report mode's 7-step approach, and fix mode's baseline → plan+apply → verify recipe ("Run `audit_diff({ audit_name: "before" })`... Confirm `-fixed` covers your targets and `+new` is empty") plus "When to bail" ("investigate and repair scoped regressions, rerun appropriate checks... Do not claim success from an incomplete or stale baseline"). Edits are validated, so the destructive/batch cap does not apply; matches the 5 anchor. | 5 / 5 |
Progressive Disclosure | A single well-sectioned ~113-line file with clear headers and no nested references, but no bundle files exist while the body names flows (`audit_live`, `audit_html`) and an `audit-live-page` prompt whose definitions live outside the package with no in-package pointers, and a fairly detailed report template sits inline. Good structure with minor organization gaps — the 4 anchor. | 4 / 5 |
Total | 18 / 20 Passed |