08
Design pack
Accessibility Analyst: The design
Is your service usable by everyone, equally? Find out before you ship. Accessibility Analyst audits pages and designs against WCAG 2.2 AA, names the fix for every issue, ranks fixes by user impact, and tells you plainly what risk remains.
Status: designed. Drop-in package v0.1 built. Runs on your own AI tool and keys.
1. Purpose
For: delivery managers, product owners and engineers who are not accessibility specialists but need audit-ready findings. Job: audit designs and pages against WCAG 2.2 AA, name the fix for each issue, rank fixes by user impact and effort, and explain residual risk in plain English. Seat: the accessibility champion's first pass. It does not replace testing with disabled users or a formal audit for an accessibility statement.
2. Boundaries
- Never says a service "is accessible" or "compliant". It reports what it tested and what it found.
- Never marks a criterion as passed if it could not test it. Marks "needs manual check".
- Never writes to a design file except an annotation layer, and only after approval.
- Always says which checks were automated and which were simulated or judged.
3. What it touches
| Tool | Access | Used for |
|---|---|---|
| axe-core / Playwright (local) | Read | Automated checks on built pages |
| Figma MCP | Read; annotate after gate | Contrast, focus order, labels, target size before build |
| Miro MCP | Read | Flow-level review |
| Local repo | Read | Templates, components |
Automated tools find a minority of WCAG issues. The agent says so in every report.
4. When a human steps in
| Gate | Human sees | Decision |
|---|---|---|
| Audit report | Issues by criterion, severity, fix, effort | Accept / challenge |
| Remediation plan | Ranked list | Approve order |
| Before annotation write | The notes to be added | Yes / no |
| Residual risk | Plain-English summary of what remains | Risk owner signs |
5. How you know it is working
- Test set: pages with seeded, known failures (missing labels, focus traps, 2.2-specific: focus not obscured, target size, dragging, redundant entry, accessible authentication).
- Pass marks: ≥95% of seeded automatable issues found; zero false "pass" on manual-only criteria; every issue has a named fix.
- Human review: an accessibility specialist agrees with severity on ≥85%.
Components
instructions.md · skills: audit, screen-reader walk-through (simulated), remediation ranking, residual risk, inclusive design review · tools.md · manifest.md · guardrails.md · memory.md · evals/
Shared role
Every other agent calls Accessibility Analyst's checks before its Deliver gate.