Workflow

The 4-Skill Lifecycle

The bundle covers the full accessibility development cycle across 8 steps. The critic runs at two checkpoints. The perspective audit is triggered by escalation — not always-on.

Escalation branches — steps 3 and 8 are conditional. When the planner or critic assigns MEDIUM or HIGH alarm to one or more perspectives, run /perspective-audit for a focused deep review of those perspectives only.

/a11y-planner step 1
/a11y-critic steps 2 & 7 (dual checkpoint)
/perspective-audit steps 3 & 8 (escalation only)
/a11y-test step 6
perspective-audit

The 7 Perspectives

Each perspective covers a distinct disability or situational access scenario. Together they map 52 WCAG 2.2 AA criteria with one primary owner per criterion.

Magnification & Reflow

Browser zoom, screen magnifier, 320px reflow

1.4.4 1.4.10 1.4.12 1.4.13 2.4.11

Environmental Contrast

High-contrast mode, outdoor brightness, color vision deficiency

1.3.3 1.4.1 1.4.3 1.4.11

Vestibular & Motion

Motion sensitivity, prefers-reduced-motion, parallax

2.2.2 2.3.1 2.5.4

Auditory Access

Captions, transcripts, visual alternatives to audio

1.2.1 1.2.2 1.2.3 1.2.5 1.4.2

Keyboard & Motor

Keyboard-only, switch access, voice control, limited dexterity

2.1.1 2.1.2 2.4.1 2.4.3 2.4.7 2.5.7 2.5.8

Screen Reader & Semantic

NVDA, JAWS, VoiceOver — semantic structure, live regions

1.1.1 1.3.1 1.3.2 4.1.2 4.1.3

Cognitive & Neurodivergent

Plain language, information density, consistent navigation

1.3.5 2.2.1 3.2.1 3.2.4 3.3.3 3.3.4 3.3.7 3.3.8
How escalation works

Standard vs. Perspective-Enhanced Review

The perspective enhancement changes the output structure, not the number of findings. The alarm level table is produced only in condition C (perspective-aware).

Standard review

Flat findings list — no perspective framing, no escalation signal.

  • Critical Carousel auto-advances without pause control. Fails WCAG 2.2.2.
  • Critical No prefers-reduced-motion media query on slide transitions.
  • Warning Slide counter text not announced to screen readers on change.
  • Warning Navigation dots lack descriptive accessible labels.

No perspective attribution. No escalation signal. User groups generic.

Perspective-enhanced review

Alarm level table — 7 perspectives, each with a level and rationale.

Perspective Alarm Rationale
Vestibular & Motion HIGH Continuous auto-advance, no motion opt-out. BPPV users affected.
Cognitive HIGH 3 s auto-advance too fast; reading speed varies widely.
Screen Reader MEDIUM Live region missing; slide count not announced on change.
Keyboard & Motor LOW Arrow key navigation present; focus visible.
Magnification LOW Reflows cleanly at 320px; no horizontal scroll.
Contrast LOW Text on slide overlay passes AA.
Auditory LOW No audio content in carousel.

Vestibular=HIGH and Cognitive=HIGH trigger /perspective-audit for those two perspectives.

Evaluation results

What the Pilot Measured

53 blind evaluations across 3 experimental rounds plus calibration fixtures. The primary quantitative metric is alarm level accuracy, not finding count.

35/35
Alarm level exact matches across 5 calibration archetypes (100%)
0%
False positive rate on clean fixtures, all rounds
53
Total blind evaluations run across 5 rounds
25+5
Fixtures — 25 main + 5 calibration, 4 categories

Calibration Results

Target was ≥80% exact match on alarm levels. Each calibration fixture spans a different component archetype from quiet (static blog, mostly LOW) to loud (animated dashboard, 4 HIGHs).

Fixture Alarm profile Exact matches Score
Static blog LOW-heavy (5L, 2M) 7/7 100%
Login / CAPTCHA 2 HIGH (Keyboard, Cognitive) 7/7 100%
Video tutorial 2 HIGH (Auditory, Screen Reader) 7/7 100%
Animated dashboard 4 HIGH (Vestibular, Contrast, Keyboard, SR) 7/7 100%
Drag-drop kanban 2 HIGH (Keyboard, Cognitive) 7/7 100%
Total Mixed LOW / MEDIUM / HIGH 35/35 100%

What the Pilot Found

Finding count did not discriminate between conditions because Claude models find code-level bugs mechanically regardless of prompting. The perspective enhancement's contribution is in output structure and escalation accuracy.

Output structure
Alarm level table + findings grouped by perspective. Standard review produces a flat list.
User group specificity
Names specific conditions (BPPV, deuteranopia) rather than generic "users with disabilities."
Escalation signal
HIGH / MEDIUM / LOW per perspective tells you which dimensions need deep review and which don't.
Fixture categories
HAS-BUGS (new dimension), HAS-BUGS (existing dimension), CLEAN, ADVERSARIAL.
Get started

Install & Commands

Install the bundle, then invoke the skills directly in Claude Code.

npx skills add zivtech/a11y-meta-skills
/a11y-planner

Design accessibility before coding. Produces a 9-phase accessibility plan including semantic structure, APG pattern choice, keyboard behavior, and testing strategy.

/a11y-critic

Review plans or implementations. Runs at two lifecycle checkpoints — after planning and after testing. Produces severity-weighted findings with a mandatory multi-perspective pass.

/a11y-test

Run real accessibility tests. Playwright keyboard interaction tests, axe-core scanning, eslint-plugin-jsx-a11y static analysis, and visual regression with screenshots.

/perspective-audit

Deep review from escalated perspectives. Triggered when the planner or critic assigns MEDIUM or HIGH alarm — runs the full JTBD checklist only for those perspectives.