Frontend: accessibility passes that finally get done

A frontend lead admitted that her team's accessibility process was "we run axe before each release and fix the highest-severity warnings." The middle and lower severity warnings accumulated. The annual a11y review revealed dozens of issues nobody had time to address.

Claude Code makes a11y a continuous discipline rather than an annual cleanup. The audit runs faster. The findings get prioritised. The fixes get scaffolded. The ritual becomes maintainable.

The combo: axe + LLM review

Axe (or its analogues) catches mechanical issues — missing alt text, label-input mismatches, contrast ratios, ARIA misuse. It produces a structured report.

The AI augments axe by:

Reading the codebase context for each finding.
Drafting a fix that respects the existing patterns.
Identifying related issues axe missed (semantic structure problems, focus management, screen-reader experience).
Prioritising findings by user impact, not just severity score.

The combo catches issues neither alone catches. Axe alone misses semantic issues. LLM alone misses mechanical ones.

Issue triage

A typical post-audit report has 50-200 findings. The AI's triage:

Block release. Critical issues (missing alt text on functional images, focus traps, keyboard inaccessibility).
Fix this sprint. High-impact issues (contrast failures, label issues).
Fix this month. Medium-impact issues (suboptimal heading structure, redundant ARIA).
Backlog or living-with. Low-impact issues (verbose ARIA, minor heading tweaks).

The triage is reviewed by a human (a11y lead or senior engineer). Some findings get reclassified. The list becomes actionable.

Fix-recipe library

Most a11y issues fall into a small number of patterns. Each pattern has a known fix. The team's fix-recipe library:

Each recipe has a name, a description of when it applies, and a concrete code example.
The library lives in the codebase, in a place engineers can find it.
The AI references the library when drafting fixes.

After two quarters of building the library, a11y findings get fixed in minutes by referring to the matching recipe. New patterns get added; the library grows.

Reviewer loop

Each a11y fix is a PR. The PR review:

Confirms the fix matches the recipe.
Tests the fix manually with keyboard and (where appropriate) a screen reader.
Confirms no regressions in nearby components.

A11y review is a learned skill. The AI helps with the mechanical checks. The engineer's understanding is what catches the subtler issues.

A monthly a11y ritual

A team that adopts the monthly ritual:

Last Tuesday of the month, the a11y agent runs a fresh audit.
The triaged report goes to the team's project tracker.
The team picks 5-10 fixes from the high-priority list to address that week.
Critical issues get fixed within 48 hours regardless.
Quarterly, the team reviews coverage and patterns.

Within a year, the team's a11y posture is qualitatively different. Annual scrambles disappear. Customer complaints about accessibility decrease. The team's skill at recognising a11y issues during normal development climbs.

What the AI cannot do

The AI can catch many issues. It cannot:

Test with real assistive technology. Screen readers behave differently across products and versions; the AI's predictions of behaviour are not authoritative.
Talk to users with disabilities. A11y is a usability concern; the test is whether real users can use the product. Synthetic testing is necessary but insufficient.
Understand cultural and contextual factors that affect a11y (cognitive load varies by language, region, education).

A team relying entirely on AI for a11y is not actually accessible. The AI is the second layer; lived testing is the first.

What we won't ship

Auto-fixing a11y issues without human review. The fix is an engineering decision.

Skipping manual screen-reader testing for critical paths.

A11y reports without action plans. Reports without action are theatre.

Generated alt text for functional images without human review. The wrong alt text is worse than missing alt text.

How to start

Run the audit on the existing codebase. Triage. Pick five high-priority fixes. Build the first five recipes. Establish the monthly ritual. By month three, the team has a working library and a sustainable rhythm.

Close

Accessibility done well is continuous, not seasonal. The audit and triage with Claude Code make continuous practical for teams that previously couldn't sustain it. The fix-recipe library compounds over quarters. The ritual builds skill. The product becomes accessible because the discipline holds.