In this article
accessibility
| Field | Value |
|---|---|
| Kind | skill |
| Source | .github/skills/accessibility/accessibility |
| Invocation | Loaded on demand by referencing agents |
| Interactive | No |
What it does
Consolidated accessibility skill entrypoint for WCAG 2.2, ARIA Authoring Practices, cognitive accessibility, Section 508, EN 301 549, design intent verification, and the Accessibility Planner workflow.
When to use it
Accessibility Planner and reviewing workflows load this skill for framework selection, standards mapping, evidence planning, scanning, and design-intent verification. It is load-only, not a slash command. Choose planning when you need to define scope and evidence; choose the scanner or runtime harness only when authorized targets and their separate prerequisites are ready.
A static scan cannot settle keyboard behavior, announcement correctness, reflow, or every semantic issue. The skill distinguishes methods that decide a criterion from those that merely inform it. Harness completion and AT-stack calibration are not conformance verdicts.
Example usage
Ask Accessibility Planner to load accessibility for a sample checkout dialog.
Supply its states, keyboard expectations, intended announcements, and target
framework/level. Request an evidence plan before any browser or scanner execution.
Expect mappings for the relevant surface/state pairs, suitable verification
methods, and manual assistive-technology work where automation cannot decide.
After separately authorizing a local test target and prerequisites, a coverage
workflow can collect probe results and render a matrix, EARL output, and manual
test plans. Success means each claimed result has an adequate method and evidence;
inconclusive observations remain cantTell, not passes. An authored design-intent
record is read-only during verification, and missing checks stay visible.
Remote scanning needs explicit host authorization, which constrains only the initial target, not every browser-derived request. Do not use confidential pages or unapproved networks, reproduce restricted standards text, or substitute automation for required qualified human review.