Agent Skills adds an accessibility skill that audits your UI for WCAG 2.2 AA

1 day ago

Addy Osmani's Agent Skills adds accessibility-engineering, a skill that audits a page you already built for WCAG 2.2 AA with pinned axe-core and pa11y, refuses to call a partial scan clean, then walks the keyboard and screen reader checks the scanners cannot do. It comes with an accessibility-auditor persona and an /a11y command for Claude Code, Gemini CLI and Antigravity. Merged November 18, 2026, in Agent Skills 0.6.17.

Ask about this presentation

Chapters

  1. 0:00Agent Skills adds an accessibility skill that audits your UI for WCAG 2.2 AA
  2. 0:20Two pinned scanners
  3. 0:33A partial run is a failure
  4. 0:44Findings sorted by severity
  5. 0:57The checks scanners cannot do
  6. 1:13Untested stays untested
  7. 1:30Touch targets drop to 24 pixels
  8. 1:45Who gets it
Show transcript
  1. Addy Osmani's Agent Skills has a new skill, accessibility engineering. It audits a page you have already built against WCAG two point two, level double A, then fixes what it finds. Your agent now runs the scanners, then the keyboard and screen reader checks they cannot do.

    Slide text
    github.com/addyosmani/agent-skills · accessibility-engineering merged Nov 18, 2026
    axe-core passed, so we’re done keyboard · screen reader · contrast
    addyosmani/agent-skills · PR #261 · skills/accessibility-engineering/SKILL.md
  2. The automated half is one script. It runs axe-core and pa11y at pinned versions, so every run behaves the same, and writes one report with each tool's version and result.

    Slide text
    github.com/addyosmani/agent-skills scripts/audit-a11y.sh

    One script runs axe-core and pa11y at pinned versions

    your page a URL or an HTML file
    →
    axe-core · pa11y fixed versions, every run
    →
    one JSON report each tool’s version, status, exit code
    addyosmani/agent-skills · PR #261 · skills/accessibility-engineering/scripts/audit-a11y.sh
  3. If axe-core or pa11y crashes or writes an unreadable report, the script stops with an error. One scanner's empty list never becomes a clean result of zero violations.

    Slide text
    github.com/addyosmani/agent-skills scripts/audit-a11y.sh

    If either scanner fails, the audit fails instead of reporting zero

    # A URL audit promises both engines.
    # Never turn a partial run into a clean result.
    "incomplete audit; refusing to report a violation total"
    exit 1
    violations: 0
    addyosmani/agent-skills · PR #261 · skills/accessibility-engineering/scripts/audit-a11y.sh
  4. Then the agent triages. Each violation goes into critical, serious, moderate or minor, and the agent inspects the element in the browser to rule out a false positive.

    Slide text
    github.com/addyosmani/agent-skills accessibility-engineering · Phase 2

    Every violation is sorted into four levels and checked by hand

    Critical blocks a screen reader: empty buttons, no alt text
    Serious low contrast, missing form labels
    Moderate skipped heading levels, redundant ARIA
    Minor missing lang on inner elements
    addyosmani/agent-skills · PR #261 · skills/accessibility-engineering/SKILL.md
  5. A page can pass axe-core and still be unusable. So the skill walks the keyboard path, then VoiceOver or NVDA by heading, landmark and link, then contrast, two hundred percent zoom and reduced motion.

    Slide text
    github.com/addyosmani/agent-skills accessibility-engineering · Phase 3

    Scanners are a smoke test; the keyboard and the screen reader finish the audit

    keyboard Tab to every control, Escape out of modals, no traps
    VoiceOver · NVDA headings, landmarks, links, form labels, live updates
    visual 4.5:1 text contrast, 200% zoom, reduced motion
    addyosmani/agent-skills · PR #261 · skills/accessibility-engineering/SKILL.md
  6. The new accessibility auditor persona ties every finding to a WCAG success criterion. A finding about what a screen reader announces stays tagged manual until someone runs that test with a named browser and screen reader.

    Slide text
    github.com/addyosmani/agent-skills agents/accessibility-auditor.md

    A screen reader result nobody tested is marked manual, never verified

    1. Map every finding to a WCAG 2.2 Success Criterion.
    2. When a finding depends on assistive-technology behavior,
       name the browser and screen reader pair to test,
       tag it `manual`, and never report it as verified
       unless that test was actually performed.
    addyosmani/agent-skills · PR #261 · agents/accessibility-auditor.md, Rules 1–2
  7. The pack's accessibility checklist used to ask for touch targets of forty-four by forty-four pixels, the stricter triple A level. It now follows WCAG two point two at double A: twenty-four by twenty-four pixels.

    Slide text
    github.com/addyosmani/agent-skills references/accessibility-checklist.md

    The shared checklist moves to WCAG 2.2 AA, so the touch target rule changes

    before · WCAG 2.1
    44×44
    px on mobile, now named the AAA target
    now · SC 2.5.8, AA
    24×24
    CSS px, with its spacing and exception rules
    addyosmani/agent-skills · PR #261 · references/accessibility-checklist.md
  8. The skill merged on November 18, with an /a11y command for Claude Code, Gemini CLI and Antigravity. If your agent ships screens nobody has audited, I would add it. Serve the page at a URL: a plain file gets only pa11y. And it audits what exists; building new UI stays with the frontend skill.

    Slide text
    github.com/addyosmani/agent-skills merged Nov 18, 2026

    Worth adding if your agent ships UI that has never had an accessibility pass

    /a11y in Claude Code, Gemini CLI and Antigravity
    URL mode runs both scanners; a plain file gets pa11y only. Building new UI stays with frontend-ui-engineering.

    addyosmani/agent-skills · PR #261 · README.md, .claude/commands/a11y.md

What happened in AI today, in your inbox

One short email each evening. Read it and you’re caught up.

Built with ❤️ in New York