Imported from OktayCopurlu/ai-shared (.github/workflows/skills/validating-ui/SKILL.md). Install upstream with npx skills add OktayCopurlu/ai-shared --skill validating-ui. Copyright stays with the author.

UI Validation

Defines what to validate in the browser after frontend implementation and how to decide pass or fail.

This skill does not explain every browser command. In gh-aw workflows, use playwright-cli from bash. It does not cover accessibility — load a11y-audit for that.

When To Use

• The change visibly affects rendered UI (component changes, layout, styling, new elements).
• After implementation is complete and quality gates (lint, types, tests) have passed.
• Before committing the final state or creating a PR.

Validation Checklist

Run these checks in order. If a critical failure blocks the target UI, run the Validation Recovery Protocol before marking anything blocked or not verified.

1. Page Load

• Page loads without a blank screen or crash.
• No unhandled errors in the console.
• No failed network requests that block rendering (4xx/5xx on critical resources).
• Hydration completes without mismatch warnings (check console for hydration errors).

2. Visual Correctness

• Test at both desktop (≥1280px) and mobile (375px) viewports. Both are mandatory unless the change is explicitly scoped to a single layout.
• The changed UI renders as expected at each viewport.
• Text content, images, and layout match what the ticket describes.
• Interactive states work: hover, focus, active, disabled, loading, empty, error.
• For multi-step flows or wizards, exercise forward and backward navigation. Verify previous selections/input are restored unless the ticket explicitly resets them, and confirm focus remains visible and is not clipped after each step transition.
• No layout shifts, overlapping elements, or clipped content in the affected area.
• Layout transitions between breakpoints do not break (no overlapping, collapsing, or hidden content).

3. Design Comparison

Only when a Figma reference exists in the ticket:

• Token-level fidelity check (required, not optional). Follow ~/.ai-shared/references/ui-fidelity-check.md — the three-way Figma / codebase / browser comparison, property list, viewport matrix, and shared-component consumer rule live there. Do not claim a UI change matches Figma without running it.
• If the design shows multiple states (empty, loading, populated, error), verify each one that is reachable at both viewports.
• If the ticket or Figma has multiple platform variants, select the web-specific design node before judging spacing or typography.
• Screenshots are supporting evidence, not the primary check. Do not approve design fidelity on screenshots alone.
• If visual polish, copy, or design intent remains subjective after the fidelity check, prepare a designer handoff with the exact preview URL or Storybook story, viewport, Figma node, and screenshots instead of claiming full design approval.

4. Runtime Health

• No new console errors or warnings introduced by the change.
• Network requests relevant to the feature return expected data (check response payloads if needed).
• GraphQL queries return the expected shape when the feature uses GraphQL.
• No infinite loops, repeated requests, or excessive re-renders visible in console or network tab.

5. Regression

• Navigate to the page or flow that existed before the change.
• Verify that existing behavior still works — especially adjacent UI that shares the same container or data source.
• If the change affects a shared component, check at least one other consumer.

6. Tracking

Only when the ticket mentions analytics or tracking:

• Verify that expected tracking events fire with correct payloads.
• Use amplitude-analytics skill or network tab to confirm event names and properties.

How To Run

1. Ensure local dev server is running.
2. Navigate, interact, and snapshot — in gh-aw workflows, use playwright-cli browser_navigate, browser_snapshot, browser_take_screenshot, and browser_run_code from bash.
3. For deeper inspection, use console messages, network requests, snapshots, and page evaluation through playwright-cli browser_run_code.
4. If a11y is relevant, load a11y-audit separately — do not mix a11y findings into this validation.
5. If the change is behind a feature flag, experiment, or A/B test, follow the experiment override protocol below. Test both enabled/disabled or control/treatment states when they are controllable.

Validation Recovery Protocol

Use this when the expected UI is not immediately reachable: localhost will not open, the preview URL is missing, the route 404s, the changed section is not visible, the PDP/product data does not show the component, a locale/market differs, auth is missing, or a flag/experiment hides the UI.

Do not mark UI validation as skipped, blocked, or not verified after a single failed attempt. First try concrete recovery paths and record what happened.

1. Capture the failed attempt: URL, viewport, console/network symptom, screenshot or DOM observation, and why it did not reach the target UI.
2. Verify the runtime path:
   • confirm the dev server command and port from repo scripts/docs
   • if localhost fails, try the PR preview when available
   • if the preview is missing, try the PR branch locally
   • if the route is unclear, search the app routes, navigation, ticket links, and changed files for the intended path
3. Verify prerequisites before giving up:
   • auth/session state
   • locale, market, country, currency, or language path
   • feature flag or experiment assignment
   • product/test data, query parameters, or CMS/API data needed to render the section
4. Try an alternate render surface when the production route cannot expose the UI:
   • direct route or deep link from source code
   • app navigation from a known working page
   • Storybook, component playground, or docs page for the changed component
   • local fixture, mock, or temporary hard-code only for validation evidence; remove it before finishing and label the evidence clearly
5. Once the target DOM is reachable, rerun the required viewport and token-level checks. Do not stop at page-load evidence if the ticket requires design fidelity.
6. Ask the user for help only after exhausting reasonable recovery paths or hitting a hard blocker such as missing credentials, unavailable environment access, or a build/runtime failure that prevents any browser surface from rendering.

A blocked/not-verified verdict must include the recovery attempts, grouped by path tried. The minimum bar is: one environment check, one route/data/flag prerequisite check, and one alternate render-surface attempt, unless a hard blocker makes one of those categories impossible.

Experiment Override Protocol

Use this when a UI change is gated by a feature flag, experiment, or A/B test.

1. Identify the exact flag or experiment key, expected group names, assignment SDK/helper, and whether assignment is server-side or client-side. Use the ticket, PR description, diff, linked docs, and nearby source code.
2. Look for supported QA override mechanisms before changing browser state: URL parameters, cookies, localStorage, sessionStorage, SDK debug APIs, preview flag endpoints, or documented browser extensions.
3. Apply only a confirmed override. Do not guess cookie or storage keys from the experiment name.
4. For cookie or storage overrides, set the value directly on the current origin with playwright-cli browser_run_code, then reload. If storage helper tools are unavailable, use page evaluation with the confirmed cookie or storage key/value.
5. Reload after applying the override and verify the active variant through rendered UI, exposure/tracking payload, network response, or runtime state. A stored value alone is not enough evidence.
6. For localhost-only validation, if the confirmed runtime override cannot be applied because assignment is server-side, extension-only, or unavailable in automation, temporarily hard-code or stub the confirmed flag/experiment return value in the local working tree. Remove the temporary change before finishing and label the evidence as local hard-code validation.
7. If the only available switch is an extension or admin UI that automation cannot operate, and local hard-code validation is not suitable, mark that variant as awaiting user-assisted override and include the exact key/group the user should select.

Verdict

After running the checklist, state one of:

• Pass: All checks passed. No issues found.
• Pass with notes: Minor issues exist but do not block the ticket. List them.
• Fail: One or more checks failed. List the failures with evidence (console output, screenshot description, or observed behavior).

Verdict Format

## UI Validation

- Viewport(s) tested: <list>
- Design reference compared: yes / no / not available
- Token-level fidelity check: pass / pass with deviations / fail / not applicable
- Tracking verified: yes / no / not applicable

### Verdict: Pass | Pass with notes | Fail

### Findings

- [check name] result — evidence or note (for token deviations: `property`: Figma `X` → computed `Y`)

### Not Verified

- what was skipped and why

Rules

• Do not validate UI before quality gates pass — fix lint/type/test errors first.
• Always test both desktop and mobile viewports. If skipping one, state the reason explicitly in the verdict.
• Do not claim design match without actually comparing to Figma.
• Do not use an app/native/mobile-platform design node as the source of truth for web UI unless the ticket explicitly says to.
• Do not treat agent-side UI validation as designer or UX approval. Report the concrete evidence and any design handoff still needed.
• Do not mix accessibility findings into this validation — use a11y-audit for that.
• If the page cannot be loaded locally, run the Validation Recovery Protocol before listing anything as blocked or not verified.
• Do not call a treatment variant blocked after arbitrary cookie attempts; first prove the supported override path or report that the variant needs user-assisted/server-side allocation.
• Do not leave temporary localhost hard-codes or stubs in the final diff unless the user explicitly asks for that implementation change.

Common Rationalizations

Red Flags

• Verdict says "Pass" but only one viewport was tested
• Design comparison claimed without a Figma link in the ticket
• Design comparison claimed without a token-level computed-style check — only screenshots or visual side-by-side
• Hardcoded numbers used when matching design-system tokens exist
• Design feedback comes from a different platform node than the implemented web surface
• Console errors dismissed as "pre-existing" without verification
• No regression check on adjacent UI
• Subjective UX or visual-polish questions are marked fully verified without designer handoff evidence
• Validation done before quality gates passed
• Tracking verification skipped when the ticket mentions analytics
• UI validation marked skipped/blocked after one failed URL, route, or section lookup
• No recovery attempts listed for a missing preview, broken localhost, hidden section, locale mismatch, or gated UI

Verification

After completing UI validation:

• Page loads without errors or blank screen
• Both desktop (≥1280px) and mobile (375px) viewports tested
• Design comparison done against Figma (if reference exists)
• Token-level computed-style check completed for each changed/new component (spacing, sizing, typography, color, borders) — or marked not applicable with reason
• Design-system tokens used instead of hardcoded values where available
• Web-specific design node was used when multiple platform variants exist
• Designer/UX handoff evidence was prepared when subjective visual approval is still needed
• No new console errors or warnings
• Regression check on adjacent UI completed
• Tracking verified (if applicable)
• Recovery protocol followed and attempts recorded when the target UI was not immediately reachable
• Verdict stated with evidence

See Also

• a11y-audit — accessibility validation (separate from UI validation)
• playwright-cli — browser automation for navigation and interaction in gh-aw workflows
• ~/.ai-shared/references/performance-checklist.md — Core Web Vitals and frontend performance checks