Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use for web accessibility work in HTML, JSX, CSS, ARIA, keyboard, forms, contrast, modals, live regions, headings, links, tables, or WCAG review; starts accessibility-lead first and uses tool_search if subagent tools are lazy-loaded.
.claude/skills/community-access-web-accessibility/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-11 | ✗→✓ | ▲ Improved | 72% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -39% | 0% |
| case-16 | ✗→✓ | ▲ Improved | 31% | 0% |
| case-17 | ✗→✓ | ▲ Improved | 15% | 0% |
| case-18 | ✗→✓ | ▲ Improved | 116% | 0% |
Use this skill for web UI accessibility work in Codex.
Installing Accessibility Agents for Codex is the user's standing request to use the Accessibility Agents subagent workflow for accessibility work. Treat every user-facing web accessibility task as authorized to spawn accessibility-lead unless the user explicitly asks for a single-agent pass.
accessibility-lead as a Codex custom subagent for every user-facing web accessibility task. Do not satisfy this step by reading the lead reference inline. The lead coordinates the same specialist team used by Claude Code.multi_agent_v1.spawn_agent is not already in the active tool list, call tool_search for multi-agent subagent accessibility before saying subagents are unavailable. Do not downgrade to local-only review just because the spawn tool was lazy-loaded.multi_agent_v1.spawn_agent, stop and ask the user to enable subagents/Accessibility Agents before continuing, unless the user explicitly authorizes a local fallback.{ "type": "skill", "name": "web-accessibility", "path": "/Users/taylorarndt/.agents/skills/web-accessibility/SKILL.md" }. Also include the relevant specialist reference path or excerpt when the task depends on a specialist workflow.codex-plugin/references/specialists/accessibility-lead.md and codex-plugin/references/specialists/index.json when available. In installed Codex plugin layouts, use .agents/plugins/a11y-agents-codex/references/specialists/ or ~/.agents/plugins/a11y-agents-codex/references/specialists/. Use the lead decision matrix and the index to select relevant specialist references and Codex subagents..a11y-agents/extensions/, ~/.a11y-agents/extensions/, and this plugin's extensions/ directory.accessibility-lead, the root session must spawn accessibility-lead and the selected specialists directly, then ask the lead to synthesize the results.accessibility-lead and every selected specialist to complete before giving the user a final answer. Do not treat a started lead as a completed review.accessibility-lead, aria-specialist, keyboard-navigator, contrast-master, forms-specialist, modal-specialist, live-region-controller, alt-text-headings, tables-data-specialist, link-checkeraccessibility-lead, aria-specialist, keyboard-navigator, alt-text-headings, plus domain specialists for forms, contrast, modals, live regions, tables, links, media, mobile, i18n, or cognitive accessibility as neededaccessibility-lead, keyboard-navigator, plus any specialists matching the diffpr-review plus any web specialists matching the diffaccessibility-lead plus the single most relevant specialist, followed by the lead final checklistaccessibility-lead, modal-specialist, keyboard-navigator, aria-specialist, and alt-text-headingsDo not expose all specialists as top-level skills. Keep the router surface small and load deep instructions lazily.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 6,917 | 4,639 | -33% | 1 | 1 | 0% | 1,206 | 1,177 | -2% | 0 | 0 | — |
case-02 | fail→fail | 3,277 | 5,436 | +66% | 1 | 1 | 0% | 524 | 1,236 | +136% | 0 | 0 | — |
case-03 | fail→fail | 15,644 | 4,841 | -69% | 1 | 1 | 0% | 1,626 | 1,276 | -22% | 0 | 0 | — |
case-04 | fail→fail | 13,822 | 4,964 | -64% | 1 | 1 | 0% | 2,294 | 1,283 | -44% | 0 | 0 | — |
case-05 | fail→fail | 9,186 | 7,671 | -16% | 1 | 1 | 0% | 1,516 | 1,620 | +7% | 0 | 0 | — |
case-06 | fail→fail | 9,009 | 5,326 | -41% | 1 | 1 | 0% | 1,305 | 1,296 | -1% | 0 | 0 | — |
case-07 | fail→fail | 10,889 | 5,684 | -48% | 1 | 1 | 0% | 2,038 | 1,351 | -34% | 0 | 0 | — |
case-08 | fail→fail | 5,166 | 9,184 | +78% | 1 | 1 | 0% | 203 | 2,087 | +928% | 0 | 0 | — |
case-09 | fail→fail | 4,057 | 7,072 | +74% | 1 | 1 | 0% | 146 | 1,335 | +814% | 0 | 0 | — |
case-10 | pass→pass | 17,928 | 8,191 | -54% | 1 | 1 | 0% | 2,709 | 2,342 | -14% | 0 | 0 | — |
case-11 | fail→pass | 19,030 | 3,573 | -81% | 1 | 1 | 0% | 922 | 1,588 | +72% | 0 | 0 | — |
case-12 | fail→pass | 11,869 | 2,509 | -79% | 1 | 1 | 0% | 2,209 | 1,338 | -39% | 0 | 0 | — |
case-13 | pass→pass | 15,502 | 10,214 | -34% | 1 | 1 | 0% | 2,640 | 2,818 | +7% | 0 | 0 | — |
case-22 | pass→fail | 9,735 | 6,046 | -38% | 1 | 1 | 0% | 1,645 | 1,241 | -25% | 0 | 0 | — |
case-14 | pass→pass | 5,467 | 2,692 | -51% | 1 | 1 | 0% | 774 | 1,349 | +74% | 0 | 0 | — |
case-15 | pass→pass | 11,354 | 2,762 | -76% | 1 | 1 | 0% | 1,755 | 1,345 | -23% | 0 | 0 | — |
case-16 | fail→pass | 11,640 | 8,757 | -25% | 1 | 1 | 0% | 1,826 | 2,397 | +31% | 0 | 0 | — |
case-17 | fail→pass | 11,134 | 5,757 | -48% | 1 | 1 | 0% | 1,674 | 1,920 | +15% | 0 | 0 | — |
case-18 | fail→pass | 8,226 | 10,855 | +32% | 1 | 1 | 0% | 1,293 | 2,793 | +116% | 0 | 0 | — |
case-19 | fail→pass | 11,191 | 2,586 | -77% | 1 | 1 | 0% | 1,890 | 1,442 | -24% | 0 | 0 | — |
case-20 | fail→fail | 10,504 | 2,005 | -81% | 1 | 1 | 0% | 1,208 | 1,196 | -1% | 0 | 0 | — |
case-21 | pass→fail | 4,574 | 6,364 | +39% | 1 | 1 | 0% | 724 | 1,316 | +82% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 10 counted toward the lift figure. The other 12 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +18 percentage points is the difference between those two pass rates over the 10 comparable cases. 2 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.