Side by side
Best for hierarchy, missing components, reflow, and reading the whole screen.
Operator handbook · 2026
Turn a visual difference into reproducible evidence.
The workbench helps a reviewer compare a baseline screenshot with a localized screenshot, inspect the same region in several views, document what is visible, and export a portable review record. It does not decide whether a translation is correct.
Before you begin: remove customer data, private messages, access tokens, and other sensitive content from screenshots. Local processing reduces exposure; it does not make sensitive source material harmless.
1. Prepare a comparable pair
- Capture the same product state, route, viewport, zoom, theme, and content fixture in the baseline and target locale.
- Use stable data. A changed timestamp, avatar, advertisement, or experiment can create noise unrelated to localization.
- Keep browser chrome out of both images unless browser UI is the subject of the review.
- Name the files so a human can identify locale, viewport, route, and state without relying on filename parsing.
When the two captures use different dimensions, treat the mismatch as context first. A pixel overlay cannot align unlike canvases reliably.
2. Load and identify the evidence
Add individual baseline and localized screenshots with the labelled file controls, or use Coverage controls to choose baseline or localized folders. Every import opens a mapping preview; an ambiguous locale is never silently accepted. Enter the context the report will need: viewport, route, build, theme, or test case. The confirmed images persist in this site’s local browser storage until removed, while an exported project file is the durable backup.
Start a new review when you change the underlying screen, not merely when you change the view mode or zoom. See Supported formats for exact image, archive, locale, matrix, and aggregate-pixel limits and split large reviews before reaching them.
3. Inspect in layers
Overlay or reveal
Best for tracing moved edges and comparing the same component directly.
Difference view
Best for locating changed pixels. It cannot explain why they changed.
Pixel loupe
Turn on Pixel loupe for clipped glyphs, diacritics, punctuation, baselines, and one-pixel collisions. Point or touch the localized image, or focus the comparison surface and move the loupe with arrow keys.
4. Write useful findings
A finding should name the visible symptom, location, conditions, user impact, and expected behaviour. Prefer “Arabic checkout total overlaps the currency label at 320 px” to “RTL broken.” Mark severity from observed impact, not from how dramatic the difference image looks.
When one pair has several distinct defects, save the primary decision and use Additional issues for separate severity, category, owner, note, reproduction context, and inspection-region links. Each issue remains distinct in HTML and CSV handoff. Record uncertainty: if an image hints at a missing translation but the source string is unavailable, say that semantic correctness requires a language review.
5. Export and hand off
Review the evidence and remove accidental personal data before export. Include enough environment detail for another person to reproduce the screen. Keep original source captures in your team’s approved storage; an exported workbench report is not a replacement for a controlled test archive.
Common questions
Does the workbench upload screenshots?
No upload or analytics endpoint is included in this static build. Files are decoded in the browser. Your browser, device management, backups, extensions, or hosting environment may still have their own policies.
Can it verify translation quality?
No. A screenshot can reveal missing, duplicated, clipped, or suspicious text, but meaning, tone, terminology, grammar, and cultural appropriateness require a qualified language reviewer.
Is every pixel difference a defect?
No. Font rasterization, caret state, animation, timestamps, antialiasing, content changes, and operating-system rendering can all create legitimate differences.
Can I use only a keyboard?
The interface is designed around labelled native controls and visible focus. Use Tab and Shift+Tab to move, and Enter or Space to activate controls. Image interpretation still requires an accessible alternative process for reviewers who cannot inspect pixels.
Does a clean comparison certify accessibility?
No. Visual review does not test semantics, screen-reader output, focus order, keyboard traps, programmatic names, live regions, or many contrast states.