Visual QA

Visual QA is a review protocol. The project adapter owns capture; the skill supplies the state matrix, independent expert lenses, refutation, synthesis, and optional verified fixes.

Inputs

Infer these when safe; otherwise ask.

• adapter: project-specific capture and test instructions in adapters/<project>.md
• tier: quick, full, or deep; default full
• mode: review or fix; default review
• scope: flows and screens; default to the adapter's declared flows
• focus: optional lens or risk to emphasize

Capability discovery

Before capture:

1. Inspect the adapter and project scripts for an existing screenshot command.
2. Inventory actually available browser, device, native-UI, and image-reading tools.
3. Prefer the project's deterministic capture harness. Do not assume Playwright, a browser MCP, or a particular agent runtime exists.
4. If capture or image inspection is unavailable, stop and name the missing capability. Never fabricate screenshots or visual findings.

This repository does not ship a universal browser driver. Start from adapters/_template.md when a project has no adapter.

Pipeline

1. Capture — render a deterministic matrix of flows, states, breakpoints, and themes; add a clearly marked exploratory pass.
2. Validate — run python3 scripts/validate_manifest.py path/to/manifest.json --check-files.
3. Review — give each selected expert only its lens, negative constraint, manifest, and images. Experts do not see one another's findings.
4. Refute — assign a skeptic to every critical/high/medium finding. The default verdict is “not proven” unless pixels or interaction evidence support it.
5. Synthesize — deduplicate, preserve dissent, and rank by severity, confidence, independent votes, and user impact.
6. Fix — only in mode: fix, with explicit authority. Give disjoint file scopes to isolated implementers and apply reviewed diffs once.
7. Re-verify — recapture affected states and compare before/after for improvement and regression.

Capture contract

The matrix should cover:

• flows: the core user journeys;
• states: default, empty, long-content, loading, error/offline, disabled, and first-run;
• breakpoints: project-specific, with narrow/mobile, medium, and wide coverage;
• themes: every supported theme;
• transitions: pending, focus, hover, pressed, open/close, and success/failure where relevant.

Pin time, timezone, locale, randomness, animation policy, and fixtures where the project permits. Deterministic fixtures should be purpose-built and non-sensitive. Exploratory screenshots are useful evidence but are not goldens.

manifest.json is an array of:

{
  "label": "checkout-error-mobile-dark",
  "path": "qa-shots/checkout-error-mobile-dark.png",
  "flow": "checkout",
  "state": "error",
  "breakpoint": 320,
  "theme": "dark"
}

Paths must be repository-relative and use forward slashes.

Finding contract

Each finding must contain:

• expert slug and lens;
• manifest label and exact visible region;
• issue and governing principle;
• pixel or interaction evidence;
• severity: critical, high, medium, low, or nit;
• concrete proposed fix;
• confidence from 0 to 1;
• any additional capture required.

Silence beats weak findings. Aesthetic preference is not an objective defect. Preserve minority opinions when they are high-confidence and clearly labeled as taste.

Severity

• critical: data loss, unintended irreversible action, unreadable core content, or sensitive information exposed in-frame
• high: blocks or badly degrades a core task, fails a required accessibility criterion, or breaks a required state
• medium: material friction, inconsistency, or ambiguous affordance
• low/nit: polish with limited user impact

Guardrails

• Isolation is mandatory during review.
• Validate capture coverage before judging design.
• Real-account or sensitive screenshots are ephemeral: keep them outside version control and remove them according to project policy.
• Never commit raw captures by default. Only synthetic goldens may be committed after human ratification.
• The human owns final taste, brand direction, and golden-image approval.
• A screenshot cannot prove DOM semantics, keyboard behavior, screen-reader output, performance, or network correctness. Route those claims to appropriate tests.

Select experts from roster.md. Keep the committee as small as the risk permits.