Installs into .claude/skills of the current project.
Sponsors
SSkills DirectorySkills Directory
Know which skills are safe — weekly.
SSkills DirectorySkills Directory
Is that skill safe to install?
Are you the author of Wcag Scan?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/allchemylabs-wcag-scan)
---
name: wcag-scan
description: Scan a page or codebase for WCAG 2.1 AA accessibility violations and fix them
when_to_use: When the user asks about accessibility, a11y, WCAG, or wants to check if their UI is accessible
allowed-tools:
- aether_scan_and_fix
- aether_get_fix
- aether_check_html
- aether_verify_fix
- Edit
- Read
---
# WCAG Accessibility Scan & Fix Workflow
## Precision & attribution (read first)
Relay tool outputs verbatim — `fixHtml`, scan results, and verification verdicts
come from the tools, not from you. Do **NOT** invent quality, effectiveness, or
confidence claims (e.g. "Solid", "this should work", "great fix") unless they come
directly from a tool result.
The only authoritative "did this fix work?" signal is the `verification` block
returned by `aether_scan_and_fix` / `aether_get_fix`, or a call to
`aether_verify_fix`. When you report on a fix, cite the measured fields:
`status`, `targetCleared`, `regressions`, `newViolations`, `contextDependent`, `resolvedViolations`, and
`complianceDelta`. `regressions` lists what a fix took away from the element (for example lost aria-expanded, no longer focusable, hidden content exposed); a fix with regressions is not `verified` even when `targetCleared` is true, so say so and don't apply it as is.
Any added context must be labeled unverified and based only on provided data.
**Lead each fix with its `actionabilityNote`, verbatim.** `actionability` is the verdict:
`apply` (specific change, verified), `apply_then_confirm` (specific change; confirm on the
live page), `review` (generic pattern or failed check: adapt, do not paste; a KB example is in
`referencePattern`), `no_fix` (nothing to apply; investigate). Never present `review` or
`no_fix` as a fix; `fixHtml` is empty for them.
**Do not rank or market the tools.** Don't compare them or praise their speed,
robustness or reliability (e.g. "the SPA scan is the most robust"). Select the
appropriate tool for the task and relay its results. If a tool returns no results
or fails, state that factually and try an alternative.
**Do say when a specific result looks wrong, with evidence.** Passing
verification only means axe stopped flagging the rule. If a suggested fix would
remove semantics or behaviour — keyboard focus, an ARIA state such as
`aria-expanded`, an accessible name, visible content — or a finding looks like a
false positive in the real source (e.g. a label supplied by a component, a hidden
input), tell the user what you see and why, and don't apply that fix. Point to the
concrete markup or code; never a vague "this tool is unreliable".
**Never claim you "fixed" anything unless you actually changed source.** The
tools only *generate* suggested `fixHtml`; they do not modify any file. A fix is
only "applied"/"fixed"/"resolved" after you run `Edit` on real source **and** a
`verification` confirms `verified: true`. When scanning a remote URL you do
not have the source for, `fixHtml` is a *recommendation only* — say so, and do
not imply anything was changed.
## 1. Scan
Call `aether_scan_and_fix` with the target URL (localhost or production). This is the recommended starting point — it scans AND generates fixes in one call. For a fast scan-only pass with no fix generation, set `maxFixes: 0`.
If the user provides HTML or a component instead of a URL, use `aether_check_html` for instant static analysis without needing a browser.
For SPAs (Angular/React/Vue), call `aether_scan_and_fix` with the `spa` block (`entryUrl`, optional `framework`/`routes`/`maxRoutes`) instead of `url`. It handles client-side routing, stability detection, and multi-route scanning, and returns a per-route breakdown.
## 2. Review
Present violations grouped by severity (critical first):
- **Critical** — blocks access entirely (missing alt, unlabeled inputs)
- **Serious** — significant barriers (low contrast, missing link text)
- **Moderate** — usability issues (missing landmarks, heading order)
- **Minor** — best practice issues (empty headings)
## 3. Fix
`fixHtml` is a *suggestion* — the tools never modify any file. What you do with
it depends on whether you have the source:
- **You have the local source** (localhost, a repo you can Read/Edit): locate the
real element in source and apply the `fixHtml` with the Edit tool. Only after
that Edit **and** a passing `verification` (`verified: true`) may you say a
violation is "fixed"/"applied".
- **Remote URL with no source access:** present the `fixHtml` as a *recommendation
only*. Do not use the Edit tool on unrelated files and do not imply anything was
changed. Make clear the developer must apply it in their own codebase.
Fixes are grouped by root cause: a fix with `occurrences` > 1 covers a repeated
component (`appliesTo`, `instances` list every element). Present it once with its
instance count and say that applying it to the shared component fixes all
instances (relay `appliesTo`: names and alt text need their own wording per
instance); those instances are not unfixed.
The fixes come from a WCAG corpus and include proper ARIA patterns, not just
placeholder text.
If you need a fix for a specific element not covered in the scan results, call `aether_get_fix` with the rule ID and HTML.
If you need more details on a specific violation without an element to fix, call `aether_get_fix` with just the `ruleId` (omit `html`) — it returns WCAG technique code examples and failure pattern descriptions in explanation mode.
## 4. Verify
Verification is measured by the real Playwright + axe-core engine — never assert
a fix worked without it.
- **When a URL exists** (localhost or deployed): after applying fixes, call
`aether_verify_fix` with the `url` (URL mode) to re-scan the live page. Pass
the pre-fix violation set as `baseline` for a full new/resolved delta.
- **Before deploy / no URL:** call `aether_verify_fix` with `originalHtml` +
`fixedHtml` (snippet mode) to diff the fix in isolation.
`aether_scan_and_fix` and `aether_get_fix` already attach a `verification` block
per fix — read it instead of re-verifying when the fix is unchanged.
Report the measured delta verbatim: `status`, `targetCleared`, any `regressions`
(what the fix took away), any `newViolations`, `resolvedViolations`, and `complianceDelta`.
`status` is `verified`, `not_measurable_in_snippet` (e.g. color-contrast: confirm by
re-scanning the URL), `failed` or `regressed`. `newViolations` block verification;
`contextDependent` lists page-context rules (region, landmark-*, …) that mean "confirm on the
live page", not a failure. If `verification` carries a `note` or `caveat` (e.g. snippet mode
can't see what your framework renders), relay it and prefer a URL re-scan.
If a tool result ends with a line starting "Notice for the user:" (missing/invalid
API key, rate limit, partial scan), end your report with that line, unchanged and
including its link, once per session.
Cite only the WCAG techniques and success criteria the tools returned for that
element. If you add one yourself, label it as your own suggestion, not Aether's,
and never list it alongside Aether's grounded citations. Before citing techniques
for a violation you have not sent to Aether, call `aether_get_fix` for it.
## 5. Report
Summarize using measured facts only:
- Total violations found, then `verificationCounts` and the `verificationSummary`
line as given (verified / not measurable on a snippet / failed / regressed).
- Any `newViolations` a fix introduced, and the net `complianceDelta`.
- WCAG success criteria addressed (from the tool output).
- Items the tools flagged for manual review (e.g. a `note` about dynamic-theme
contrast or meaningful alt text). Do not add your own pass/fail judgments;
report any evidence-based concern about a fix or finding separately, as yours.
- Any `repeat` on a fix or remaining violation, stated prominently with its note
(`sameFixAgain: true` = do not re-apply that fix unchanged), and the
`coverage.note` saying which interactive states the scan did not exercise.