web-accessibility（网络可访问性）

Web Accessibility

Use this skill when the job is to name the failing accessibility surface, fix it at the right layer, and leave behind a short remediation + verification packet.

The job is not to dump generic WCAG prose, paste raw scanner output, or absorb every neighboring frontend concern.

This skill should:

1. classify the failing accessibility surface,
2. separate automated findings from manual or assistive-technology follow-up,
3. choose the smallest credible remediation packet,
4. keep routed-app feedback and focus behavior explicit,
5. route broader design/layout/component-governance work honestly.

Read these support docs first:

• references/intake-packets-and-route-outs.md
• references/audit-remediation-checklist.md
• references/handoff-boundaries.md

When to use this skill

• A team says “make this accessible” and the missing accessibility surface is still unclear
• axe, Lighthouse, Accessibility Insights, Pa11y, QA notes, or user reports need to be turned into one remediation plan instead of a raw finding dump
• You need to decide whether the main issue is semantics, keyboard/focus, labels/announcements, reflow/contrast/motion, media alternatives, or routed-app navigation feedback
• A modal, form, nav, menu, table, dashboard, or routed flow needs manual-vs-automated verification guidance before release
• Screen-reader, keyboard-only, or zoom/reflow issues are suspected, but ownership and next checks are unclear

When not to use this skill

• The main task is broad UI polish, heuristic review, hierarchy, or interface consistency → web-design-guidelines
• The main task is reusable primitive / slot / variant API design or component-family ownership → ui-component-patterns
• The main task is viewport adaptation, breakpoint strategy, container-query planning, or layout overflow control → responsive-design
• The main task is token governance, contribution rules, or system-wide frontend standards → design-system
• The task is only to run one tool and paste raw results; run the tool directly, then return here to prioritize remediation and manual follow-up

Instructions

Step 1: Frame the accessibility job before naming fixes

Capture the minimum packet first.

accessibility_intake:
  surface: page | routed-flow | modal | form | nav | menu | table | dashboard | component-family | mixed | unknown
  workflow_type: audit-review | remediation-pass | release-readiness | regression-follow-up | user-reported-bug | design-system-follow-up
  signal_source: axe | lighthouse | accessibility-insights | pa11y | manual-qa | screen-reader-report | user-feedback | mixed | unknown
  primary_surface: semantics | keyboard-focus | labels-announcements | visual-perception-reflow | media-alternatives | routed-navigation-feedback | mixed | unknown
  severity: blocker | major | medium | low | unknown
  ownership: app-team | frontend-platform | design-system | qa-accessibility | shared | unknown

Rule: do not start with “add ARIA,” “run axe again,” or “quote WCAG 2.1.” First label the failing accessibility surface.

Step 2: Choose exactly one primary remediation packet

Use the router in references/intake-packets-and-route-outs.md.

Primary packets:

1. semantics-structure
2. keyboard-focus
3. labels-announcements
4. visual-perception-reflow
5. media-alternatives
6. routed-navigation-feedback

Pick the highest-risk user-facing failure as primary. List everything else as follow-up, not as equal co-owners.

Step 3: Keep the invariants visible

These rules survive every answer:

• automated tools accelerate discovery, but they do not prove accessibility completeness
• native HTML semantics beat ARIA patching when native controls already solve the problem
• keyboard/focus failures in interactive flows usually outrank cosmetic cleanups
• routed apps need explicit focus and announcement decisions because browser navigation cues disappear
• labels, instructions, errors, and async status text are part of task completion, not polish
• every remediation recommendation needs a matching manual verification step

Step 4: Build the accessibility remediation packet

Return this structure:

# Accessibility Remediation Packet

## Scope
- Surface:
- Workflow type:
- Primary packet:
- Confidence: high | medium | low

## Current signal
- Tool or report source:
- What is already known:
- What still needs direct verification:

## Highest-risk gaps
1. ...
2. ...
3. ...

## Recommended first slice
1. ...
2. ...
3. ...

## Verification plan
- Automated re-check:
- Manual keyboard / focus:
- Manual screen-reader / AT:
- Visual / zoom / motion / reflow:

## Ownership and route-outs
- Primary owner:
- Adjacent skills / teams:

Step 5: Use the packet, not a giant checklist

Pull the packet from references/intake-packets-and-route-outs.md.

Packet rules:

• semantics-structure → landmarks, headings, lists, tables, buttons/links, form markup, native-first fixes
• keyboard-focus → tab order, visible focus, trap/escape behavior, initial focus, focus return, composite widget behavior
• labels-announcements → accessible names, helper/error relationships, live regions, status/error messaging, required-state clarity
• visual-perception-reflow → contrast, reduced motion, zoom, reflow, target visibility, state visibility
• media-alternatives → alt text, decorative-vs-informative imagery, captions, transcripts, fallback descriptions
• routed-navigation-feedback → page title / heading announcement, route-change focus strategy, SPA/app-router cues, async load state announcements

Step 6: Separate automated findings from manual proof

Use this split in every serious answer:

• Automated-friendly — missing labels, duplicate IDs, invalid ARIA, simple contrast failures, landmark/heading gaps
• Manual keyboard/focus — tab sequence, visible focus, trap behavior, initial focus, focus return, menu/listbox/dialog interaction
• Manual AT / content quality — announcement usefulness, alt-text quality, label clarity, route-change comprehension, async feedback quality

If the request starts from a scanner, say explicitly that the scanner is the input artifact, not the finished answer.

Step 7: Route adjacent work explicitly

Use these route-outs when the problem crosses boundaries:

Output expectations

A strong answer from this skill should:

1. identify the primary failing accessibility surface,
2. recommend one bounded remediation packet,
3. name the manual verification still required,
4. avoid equating a scanner pass with completion,
5. route broader frontend ownership questions outward instead of absorbing them.

Examples

Example 1: modal audit after axe

Input

Audit this checkout modal for accessibility issues and tell me what to test manually after axe.

Output direction

• choose keyboard-focus or labels-announcements
• treat axe as the input artifact, not the whole workflow
• verify initial focus, trap behavior, focus return, and announcement quality
• keep broader visual-polish critique out unless it materially affects accessibility

Example 2: routed app navigation confusion

Input

Screen-reader users say our React app changes routes but the new page is not announced and focus stays on the old nav link.

Output direction

• choose routed-navigation-feedback
• classify title/heading announcement plus focus-reset strategy
• include manual AT verification after the fix
• avoid turning the task into generic routing architecture work

Example 3: component-boundary confusion

Input

Our button primitive API is messy and some variants are not accessible. Is this web-accessibility or ui-component-patterns?

Output direction

• keep accessible-name, focus-state, and disabled/loading announcement fixes in web-accessibility
• route primitive API / variant ownership to ui-component-patterns
• split the plan instead of forcing one skill to own everything

Example 4: manual checks after Lighthouse

Input

Lighthouse says contrast is okay now. What should I still test manually?

Output direction

• treat this as verification planning, not score celebration
• verify keyboard order, focus visibility, routed-flow cues, and screen-reader announcement quality where relevant
• include zoom/reflow and reduced-motion checks when the interface depends on them

Best practices

1. Start with the user task and failing surface, not the rule number.
2. Treat automated tools as accelerators, not as proof of accessibility completeness.
3. Prefer native HTML semantics before ARIA-heavy patches.
4. Keep focus behavior intentional for overlays, route changes, and async validation.
5. Separate remediation packets from broader design-review or component-architecture work.
6. When unsure, name the manual verification still required instead of pretending the scan is enough.

References

• W3C: Evaluating Web Accessibility Overview
• W3C: Easy Checks – A First Review of Web Accessibility
• web.dev: Manual accessibility testing
• MDN: Keyboard accessible
• React Router: Accessibility
• Next.js: Accessibility