# Review Accessibility

> Review existing interface code, diffs, or rendered flows for accessibility barriers. Use only when explicitly invoked as `$review-accessibility` or `/review-accessibility`. Report findings only; do not modify code unless separately asked.

- Skill: `imkohenauser/review-accessibility` (Agent Skill, multi-file: 3 files)
- Install (CLI): `npx skillmds@latest add imkohenauser/review-accessibility`
- Raw SKILL.md: https://api.skillmd.com/api/skills/imkohenauser/review-accessibility/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Web & Frontend
- License: MIT
- Author: imkohenauser (https://skillmd.com/u/imkohenauser)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/imkohenauser/review-accessibility

---


# Review Accessibility

Review only the interface and user flow in scope. Do not edit files, implement fixes, or expand into general UI review unless the user separately asks.

Prefer the project's stated accessibility target. When none is stated, use WCAG 2.2 Level AA as the review baseline and identify checks that require manual or assistive-technology verification.

Read [references/review-checks.md](references/review-checks.md) completely before reviewing.

## Review method

1. Identify the user tasks, interactive elements, and dynamic states in scope.
2. Inspect semantics, accessible names, roles, relationships, and exposed states.
3. Inspect keyboard operation, focus order, focus visibility, focus movement, and restoration.
4. Inspect forms, errors, status messages, and other dynamic updates.
5. Inspect images, media, target sizes, color-independent cues, zoom, reflow, and motion preferences.
6. When a browser is available, test keyboard-only operation, the accessibility tree, zoom or narrow reflow, and reduced-motion settings.
7. Separate confirmed findings from checks that could not be performed.

Automated audits supplement manual review. They do not establish conformance.

## Finding requirements

Report only issues supported by code or observed behavior. Each finding must include severity, location or affected flow, current behavior, affected users and task impact, recommended outcome, and verification status.

Classify each finding as a requirement failure or a recommended improvement. For a requirement failure, cite the applicable WCAG success criterion and level or the explicit project requirement. Do not treat an advisory pattern or a criterion above the review target as a requirement unless the project adopts it.

Do not report conventions as failures without concrete user impact:

- Do not report heading-level preferences as standalone failures.
- Check target-size exceptions before reporting an undersized control.
- Do not assume a custom focus color or contrast pair passes without measuring the rendered result.
- Do not claim screen-reader behavior was verified when only source code was inspected.
- Prefer valid native behavior over a custom ARIA reconstruction.

## Severity and verdict

- **HIGH** — blocks a task, hides essential content, traps or loses focus, or makes an essential control unavailable to keyboard or assistive technology.
- **MEDIUM** — creates substantial friction, ambiguity, or unreliable operation without fully blocking the task.
- **LOW** — isolated accessibility polish with limited task impact.

List findings first, ordered by severity:

| Severity | Location | Finding and basis | User impact | Recommendation | Verification |
| --- | --- | --- | --- | --- | --- |

Use `path:line` for code findings and a concise flow or element name for runtime findings.

Close with one verdict, in this order:

- `Block` when a confirmed HIGH finding remains, even if other checks are incomplete.
- `Inconclusive` when no HIGH finding is confirmed but checks needed to judge the in-scope tasks could not be performed. Name the missing checks and what would resolve them.
- `Approve` when no HIGH finding remains and the checks needed for the scoped review are complete. Retain any MEDIUM or LOW findings.

State the inspected scope and verification limits. Mark individual unavailable checks as `Not verified`; do not imply that source inspection verifies browser or assistive-technology behavior. Approval applies only to the reviewed scope and is not a claim of complete WCAG conformance.