UI Review

Automated visual QA: interview the user, then test a web app's UI across viewports using Agent Browser, evaluate design quality, and iterate until satisfied.

Process Overview

1. Interview — Gather route, credentials, and focus areas from user
2. Login & Navigate — Open the app and authenticate
3. Multi-Viewport Capture — Screenshot at 8+ viewport sizes
4. Evaluate — Score UI against design and responsive criteria
5. Report — Deliver findings with actionable recommendations
6. Iterate — Fix issues and re-test until all checks pass

Step 1: Interview

Use AskUserQuestion to collect (skip any the user already provided):

1. App URL — Base URL (e.g., http://localhost:3000)
2. Route to test — Specific path (e.g., /dashboard, /settings/profile)
3. Test credentials — Email/username and password for login
4. Login route — Where the login form lives (default: /login or /sign-in)
5. Focus areas — What matters most: scroll behavior, layout, typography, specific components, mobile experience, etc.
6. Viewport selection — Standard set (8 viewports) or custom. See references/viewports.md.
7. Artifacts folder — An accessible path to store screenshots (default: ./ui-review-artifacts)

Keep it to 2 questions per round. Move on once you have URL, route, and credentials at minimum.

Step 2: Login & Navigate

Use Agent Browser with a named session so auth persists.

# Start session and navigate to login
agent-browser --session ui-review open "<app_url><login_route>"
agent-browser --session ui-review snapshot -i

# Fill login form (adapt selectors to actual form)
agent-browser --session ui-review fill "<email_field>" "<email>"
agent-browser --session ui-review fill "<password_field>" "<password>"
agent-browser --session ui-review click "<submit_button>"

# Wait for redirect
agent-browser --session ui-review wait --load networkidle
agent-browser --session ui-review snapshot -i

After login, navigate to the target route:

agent-browser --session ui-review open "<app_url><target_route>"
agent-browser --session ui-review wait --load networkidle

Step 3: Multi-Viewport Capture

Before capturing, set up an accessible artifacts folder with a stable layout:

export ARTIFACTS_DIR="<path>"   # default: ./ui-review-artifacts
mkdir -p "$ARTIFACTS_DIR"/{before,after,combined}

Use before/ for the initial baseline screenshots. After fixes, recapture into after/ using the same filenames to allow pairing.

Minimum viewport set — test ALL of these:

Label	Width	Height
Mobile S	375	667
Mobile L	430	932
Tablet	768	1024
Tablet L	1024	1366
Laptop	1366	768
Desktop FHD	1920	1080
Short desktop	1920	600
Narrow desktop	1024	768

For each viewport, open a separate session (most reliable for viewport sizing):

agent-browser --session "vp-<label>" open "<url>"
agent-browser --session "vp-<label>" eval "
  Object.defineProperty(window, 'innerWidth', {value: <width>, configurable: true});
  Object.defineProperty(window, 'innerHeight', {value: <height>, configurable: true});
  window.dispatchEvent(new Event('resize'));
"
agent-browser --session "vp-<label>" wait 1000
agent-browser --session "vp-<label>" screenshot "$ARTIFACTS_DIR/before/<label>.png" --full

If cookies don't transfer between sessions, log in again per session or reuse the main ui-review session and resize with eval.

Also test scroll behavior at each viewport:

agent-browser --session "vp-<label>" scroll down 500
agent-browser --session "vp-<label>" screenshot "$ARTIFACTS_DIR/before/<label>-scrolled.png"
agent-browser --session "vp-<label>" scroll up 500

Use subagents (Task tool, subagent_type=Bash) to parallelize viewport captures.

See references/viewports.md for additional viewport presets including unusual sizes.

After you have both before/ and after/, create combined images with labels:

for f in "$ARTIFACTS_DIR"/before/*.png; do
  base="$(basename "$f")"
  before="$ARTIFACTS_DIR/before/$base"
  after="$ARTIFACTS_DIR/after/$base"
  [ -f "$after" ] || continue
  magick \
    \( "$before" -background white -gravity north -splice 0x60 -pointsize 32 -fill "#111" -annotate +0+10 "BEFORE" \) \
    \( "$after" -background white -gravity north -splice 0x60 -pointsize 32 -fill "#111" -annotate +0+10 "AFTER" \) \
    +append "$ARTIFACTS_DIR/combined/$base"
done

If ImageMagick is unavailable, use an equivalent image tool to place BEFORE/AFTER labels and stitch left/right.

Step 4: Evaluate

For each screenshot, use the Read tool to view the image and evaluate against references/review-criteria.md.

Use subagents (Task tool) to parallelize evaluation — one subagent per viewport or per group of viewports.

Evaluate each screenshot for:

• Layout & Spacing — Content constrained, no overflow, proper whitespace
• Typography — Readable, proper hierarchy, comfortable line length
• Navigation & Interactive — Touch targets >=44px, button visibility
• Visual Consistency — Colors, images, decorative elements consistent
• Scroll Behavior — Smooth scroll, fixed elements don't overlap
• Breakpoint Transitions — Clean layout transitions between sizes
• Content Priority — Important content above fold

Apply $frontend-design principles:

• Does the design feel polished and intentional?
• Is there a clear visual hierarchy?
• Are colors, spacing, and typography harmonious?
• Does it avoid generic/template aesthetics?

Apply responsive design principles:

• Do breakpoint transitions feel natural?
• Is mobile content prioritized correctly?
• Are touch targets appropriately sized?
• Does the layout use available space well at each size?

Step 5: Report

Write a structured report to ui-review-report.md (or user-specified path).

Include:

• Screenshot references for each viewport
• Combined before/after images from combined/
• Per-viewport scores using criteria from references/review-criteria.md
• Cross-viewport summary
• Prioritized issue list
• Specific CSS/component fix recommendations

Step 6: Iterate

If Major or Critical issues exist:

1. Present findings with specific fix suggestions
2. If user approves, implement the fixes
3. Re-run viewport capture and evaluation (save into after/)
4. Repeat until all viewports score Pass or Minor only

Use subagents for re-test cycles to keep context clean.

Key Principles

• Named sessions — --session ui-review to persist login state
• Screenshot everything — visual evidence for every finding
• Test scroll — scrolled screenshots, not just top-of-page
• Test extremes — very short heights, very narrow widths catch edge cases
• Subagents — parallelize captures and evaluations via Task tool
• Iterate — re-test after fixes, don't stop at first report