[DT-4049] add a11y skills, update CLAUDE.md

[ardiewen]

Description & motivation 💭

Adds three accessibility skills to .claude/skills/ and a short pointer in CLAUDE.md. The skills codify the conventions and review rubrics for the A11y PR Triage workflow that landed in #3465 so they survive across sessions, contributors, and team membership changes.

• a11y-pr-review — reviewer-facing. Decodes the a11y:* labels the triage workflow applies (buckets 1–4, failure-modes, SC labels) and prescribes the review depth for each.
• a11y-manifest-discipline — author-facing. The rule for keeping scripts/a11y/manifest.yml in sync when shipping or re-bucketing a fix doc, plus the bucket framework recap and backfill commands.
• wcag-audit — auditor-facing. The in-repo distillation of the four verification work shapes, the static-vs-manual decision discipline, inventory practice, verdict logic, severity rubric, and anti-patterns the audit team has hit. The full source artefact lives in the audit team's private workspace; this is the public, auditable version.

The CLAUDE.md addition is intentionally minimal — a single paragraph pointing at the three skills. Earlier drafts inlined more guidance; on review of #3465, that pattern was flagged as the wrong place for it. Skills are the right home for procedural knowledge; CLAUDE.md should just announce that the system exists.

Screenshots (if applicable) 📸

N/A — documentation / agent-tooling change. No UI surfaces.

Design Considerations 🎨

• Per-repo divergence is intentional. This PR ships an OSS-safe variant of the same three skills also being added to temporalio/cloud-ui. The two variants drift in places — slug conventions, scope enums, the example trailer — because OSS contributors don't have visibility into private tickets, the audit team's private workspace, or the existence of a second repo. Anywhere the two had to differ, the OSS version stays explicit and the private version carries the cross-repo coordination detail.
• The wcag-audit skill is a port, not a replacement. The audit team's process artefact has worked examples that cite specific issue docs (private). That doc is the source of truth; this file is the durable, in-repo distillation that survives team changes. The Status note at the top makes this explicit, including the four criterion rows it has actually been exercised against.
• No code changes. The triage workflow, the manifest, and the validator already shipped in [DT-4048] Add accessibility PR triage and notification helpers #3465. This PR is purely the human-and-agent-readable layer on top.

Testing 🧪

How was this tested 👻

• Manual testing
• E2E tests added
• Unit tests added

Specifically:

• Read each skill end-to-end against its intended audience question ("what does a11y:broken-ref mean?", "how do I add a manifest entry?", "what's the shape of this row's verification work?") to confirm the skill answers it without needing the source artefact.
• Verified the skills surface in the Claude Code skill list (user-invocable: true frontmatter on all three) and that each one's description is specific enough to disambiguate between them.
• Cross-checked all internal references: [[wcag-audit]], [[a11y-manifest-discipline]], [[a11y-pr-review]] links resolve to sibling skill slugs.

Steps for others to test 🚶🏽‍♂️🚶🏽‍♀️

In a Claude Code session in this repo:

1. Run /a11y-pr-review and ask it to walk a labeled PR — confirm the bucket-to-review-depth mapping is what you'd expect.
2. Run /a11y-manifest-discipline and ask it to draft an entry for a hypothetical new fix doc — confirm it asks the right field questions and produces a valid YAML row.
3. Run /wcag-audit and ask it which of the four verification work shapes applies to a given matrix row — confirm the trigger criteria are clear enough to choose without flipping a coin.

Checklists

Draft Checklist

• Sanitized for OSS — no JIRA links, no cross-repo references, no audit-team-private resource names.
• CLAUDE.md shortened to a pointer per review of [DT-4048] Add accessibility PR triage and notification helpers #3465.

Merge Checklist

• Skills are discoverable (user-invocable: true).
• Cross-skill references resolve.
• CLAUDE.md addition fits the existing structure.

Issue(s) closed

N/A — internal initiative. Tracked in temporalio/cloud-ui's issue tracker.

Docs

Any docs updates needed?

None. The skills are the docs for this surface; CLAUDE.md points at them.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
holocene	Ready	Preview, Comment	May 29, 2026 6:27pm