agentleFS
AdilSiddiquiHQAdilSiddiquiHQ/a11y-skills/skills/triage-automated-results/SKILL.md

triage-automated-results

Skill0 starsChanged 30 days agoMITView on GitHub

Use when an automated accessibility scan reports issues to triage, when a Lighthouse or axe score is being treated as conformance, or when a tool finding needs a verdict, an owner, and a criterion.

The file

We split this skill into the rules it gives an agent

1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Sort every finding into three buckets: Do not "fix" a finding until it is in one:

    line 32

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Triage automated results

Criteria: SC 1.4.3 Contrast (Minimum), SC 1.1.1 Non-text Content, SC 4.1.2 Name, Role, Value, SC 1.4.11 Non-text Contrast

Automated testing finds on the order of a third of real accessibility issues. Clearing the report is not conformance, and a green score is not a claim anyone should sign. Triage is the work of turning each finding into a verdict — real, false positive, or needs a person — plus the criterion, the user impact, and the fix. The findings a tool cannot decide are where the actual problems live.

When to use

  • "Lighthouse says 96 — are we done?"
  • A CI job emits axe, WAVE, Pa11y, or IBM Equal Access results.
  • Deciding whether a finding is real, a tool limitation, or needs human judgement.
  • Someone wants to close a ticket because a scanner turned green.
  • A tool flags a rule and you need to know which criterion and what to check next.

Procedure

  1. State the limit before you start. Automated tooling checks a subset of criteria that can be decided mechanically. It cannot test keyboard operability, reading order, focus order, captions, error identification, or whether text is meaningful. Record which criteria were covered automatically and which still need a person.
  2. Sort every finding into three buckets. Do not "fix" a finding until it is in one:

    BucketMeaningExample
    confirmeda real failure against a criteriona button with no accessible name (SC 4.1.2)
    false positivea tool limitation, not a defectcontrast flagged on text over a decorative image with a solid backing
    needs humanthe tool cannot decidean alt that is present but does not describe the content (SC 1.1.1)

    The tool's "needs review" or "incomplete" bucket always lands in needs human. Those items are not passes.

  3. Read each finding as a claim about a criterion, and test the claim.

    • SC 1.4.3 Contrast (Minimum) — 4.5:1 body, 3:1 large. A checker can only compute a ratio when the background is a solid, known value. It misses text over images, gradients and opacity, text on pseudo-elements, and most state changes. Sample the real background behind the text.
    • SC 1.1.1 Non-text Content — a tool finds a missing alt. It cannot judge whether the alt describes the content. alt="image", alt="logo.png" and alt="chart" all pass the scanner and fail the criterion.
    • SC 4.1.2 Name, Role, Value — tools catch some missing names and roles. They do not catch a wrong name, a name that contradicts the visible label, or a state that is never exposed.
    • SC 1.4.11 Non-text Contrast — poorly covered, and almost never checked across hover, focus and disabled states. Check each state against 3:1 by hand.
  4. Then run the human half. Keyboard-only operation, a screen reader pass, zoom and reflow, and reading order. None of these are decided by a scanner; see the screen-reader-test-plan skill.
  5. Fix by impact, not by severity label. A tool may rate an unlabelled input as "minor"; if it blocks a user from completing a task it outranks a cosmetic "serious" finding. Name the user impact in the ticket.
  6. Re-verify the specific failure by hand. Re-run the one rule to confirm it cleared, then confirm the user-facing behaviour, rather than re-running the whole suite and reading the score.
  7. Handle findings aimed at a criterion that no longer exists. Some tools still tag duplicate-id and empty-heading findings to the removed Parsing criterion (4.1.1), which was removed in WCAG 2.2. Record those as robustness notes; they are not WCAG 2.2 failures.
<!-- An automated check is green. A human check is not.
     The image-alt rule passes because alt is present.
     SC 1.1.1 fails because the alt does not describe the content. -->
<img src="revenue-2026.png" alt="chart">

<!-- The corrected version: say what the chart shows. -->
<img src="revenue-2026.png"
     alt="Revenue rose from £2.1m in January to £3.4m in June 2026.">
// Before writing tickets, split the tool's output into the three buckets.
const cannotResolveBackground = (node) =>
  node.any?.some((c) => /background/i.test(c.message ?? '')) ?? false;

const bucket = (violations) => violations.flatMap((rule) =>
  rule.nodes.map((node) => ({
    rule: rule.id,
    target: node.target.join(' '),
    // Contrast over an image, gradient or opacity cannot be trusted blind.
    verdict: rule.id === 'color-contrast' && cannotResolveBackground(node)
      ? 'needs-human'
      : 'confirmed',
  }))
);

Verify

For each criterion, know exactly what the tool did not check:

Tool findingCriterionNot checked by the toolManual check
color-contrast1.4.3the real background behind the text (image, gradient, state)sample the pixel behind the text and compute the ratio
image-alt1.1.1whether the alt describes the contentread it and ask whether it conveys the meaning
button-name4.1.2whether the name matches the visible label and whether the state is exposedlisten to the control and operate it
(rare) non-text contrast1.4.11hover, focus, disabled and selected statesmeasure 3:1 in each state
// A rule passing is evidence about the rule, not about the user. Assert the
// behaviour the user needs, which is what the rule approximated.
const toggle = document.querySelector('#menu-button');
console.assert(toggle?.getAttribute('aria-expanded') != null,
  'expanded state is not exposed (SC 4.1.2)');
const label = toggle && document.querySelector(`#${toggle.id}`)?.textContent?.trim();
console.assert(label, 'visible label is empty (SC 2.5.3)');

Failure modes

  • Closing a ticket because the scanner is green. Automated coverage is about a third of real issues; the other two thirds are exactly the criteria that matter to users (SC 3.3.1, SC 2.4.3, SC 1.2.2, and so on).
  • Dismissing the "needs review" bucket. It is the tool telling you a human must look, not that the item is fine.
  • Accepting alt="image" because the rule passed. The name exists and says nothing, which is the failure SC 1.1.1 is about.
  • Adding aria-label to silence button-name on a decorative icon. The correct fix is usually aria-hidden="true", and the label creates a duplicate or confusing announcement.
  • Trusting an automated contrast pass on a hero image or gradient. The checker could not resolve the background, so the ratio it reported is not the ratio the user sees (SC 1.4.3).
  • Skipping non-text contrast entirely (SC 1.4.11) because no rule flags it. Fields, icons and focus indicators still need 3:1 against their surroundings.
  • Prioritising by the tool's severity label. "Minor" findings that block a task are more urgent than "serious" findings that are cosmetic.
  • Overcorrection — deleting images or hiding content from assistive tech to reach a clean score. A passing report achieved by removing content is worse than the findings it replaced.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.