bitovi/ai-enablement-prompts

Skill: Visual Diff

Compare two URLs visually — a baseline (e.g. production site) and a current (e.g. local dev server or Storybook story) — to find pixel-level differences and guide corrections.

---

When to Use

• Replicating an existing page or section in Storybook or the Astro site
• Verifying visual accuracy of a component against a reference design
• Checking responsive fidelity across breakpoints after style changes
• Iterating on CSS/HTML until a component matches its reference

When NOT to Use

• Content-only changes (text updates, copy edits) where layout is unchanged
• Comparing designs from Figma — use the Figma MCP skill instead
• Performance or accessibility audits — use Lighthouse or the a11y addon

---

Prerequisites

Before starting, confirm:

1. Playwright MCP is configured in .vscode/mcp.json with --caps=vision
2. pixelmatch and pngjs are installed: check package.json devDependencies
3. The baseline URL is accessible (external site, or a running server)
4. The current URL is accessible (dev server on :4321, Storybook on :6006, etc.)
5. The temp/ directory exists in the workspace root (it's gitignored)

If the dev server or Storybook isn't running, start them using the VS Code tasks:

• Dev server: use the Dev Server task
• Storybook: use the Storybook task

---

Workflow

Step 1: Receive inputs

The agent needs:

Input	Required	Example
Baseline URL	Yes	https://www.bitovi.com/
Current URL	Yes	http://localhost:4321/ or http://localhost:6006/?path=/story/card--default
CSS selector	No	#hero-section or .card-grid (for element-scoped comparison)
Breakpoints	No	Defaults to mobile (375×667), tablet (768×1024), desktop (1280×720)

Step 2: Screenshot the baseline

Component-scoped comparison (preferred when comparing a single component)

When comparing a specific component — especially when the Storybook story renders only that component but the baseline page has a full page around it — crop to the component's bounding box so the diff isn't polluted by unrelated page sections (nav bars, sticky overlays, other modules above/below).

How to crop on the baseline page:

1. Navigate to the baseline URL
2. Resize viewport to { width: 1280, height: 720 }
3. Scroll the target element into view:

   () => document.querySelector('<selector>').scrollIntoView({ behavior: 'instant', block: 'start' })
   

4. Measure the component's rendered height:

   () => document.querySelector('<selector>').getBoundingClientRect().height
   

5. Resize the viewport height to match that exact height (keeps width constant):

   mcp_playwright_browser_resize → { width: 1280, height: <measured height> }
   

6. Take the screenshot — it now captures only the component area.

Background bleed warning: What's visible behind the component will differ between Storybook (blank) and the baseline page (other modules). Common cases: sticky/local nav bars overlapping the top, modals/overlays with other page content visible behind them, and partially transparent components. These residual diffs cluster outside the component's own content and are not actionable — ignore them.

How to crop on the Storybook story:

Storybook iframe.html renders only the component with no surrounding chrome — no resizing needed. Just navigate and screenshot:

mcp_playwright_browser_navigate → { url: "http://localhost:6006/iframe.html?id=<story-id>&viewMode=story" }
mcp_playwright_browser_resize → { width: 1280, height: <same height as baseline crop> }
mcp_playwright_browser_take_screenshot → { filename: "temp/vdiff-current-desktop.png" }

Full-page viewport screenshot (when comparing full pages)

1. mcp_playwright_browser_navigate → { url: "<baseline-url>" }
2. mcp_playwright_browser_resize → { width: 1280, height: 720 }
3. mcp_playwright_browser_wait_for → { time: 2 }       # let page fully render
4. mcp_playwright_browser_take_screenshot → {
     type: "png",
     filename: "temp/vdiff-baseline-desktop.png"
   }

Repeat for each breakpoint:

Breakpoint	Width	Height	Filename suffix
Mobile	375	667	-mobile
Tablet	768	1024	-tablet
Desktop	1280	720	-desktop

Step 3: Screenshot the current

Repeat the exact same process for the current URL, using temp/vdiff-current-{breakpoint}.png filenames.

Important: Use the same viewport dimensions and the same cropping strategy (component-scoped or full-page) on both sides so the comparison is apples-to-apples.

Step 4: Run the diff script

For each breakpoint, run the visual-diff script in the terminal:

node scripts/visual-diff.mjs \
  --baseline temp/vdiff-baseline-desktop.png \
  --current temp/vdiff-current-desktop.png \
  --output temp/vdiff-diff-desktop.png

The script outputs JSON to stdout:

{
  "totalPixels": 921600,
  "diffPixels": 4521,
  "diffPercent": 0.49,
  "width": 1280,
  "height": 720,
  "baselineDimensions": { "width": 1280, "height": 720 },
  "currentDimensions": { "width": 1280, "height": 720 },
  "dimensionsMismatch": false,
  "diffImagePath": "/absolute/path/to/temp/vdiff-diff-desktop.png"
}

If dimensionsMismatch is true, the images had different sizes. The script pads the smaller image with transparent pixels before comparing. This often indicates a layout issue worth investigating.

Step 5: View the diff image

Start a temporary HTTP server to serve the temp/ directory, then navigate Playwright to the diff image:

# Start server in background (run once per session)
npx -y http-server temp/ -p 8787 --cors -c-1 &

Then navigate Playwright to the diff image:

mcp_playwright_browser_navigate → { url: "http://localhost:8787/vdiff-diff-desktop.png" }
mcp_playwright_browser_take_screenshot → {}   # triggers vision — agent can now see the diff

The --caps=vision flag on Playwright MCP means the agent can see the image directly in the screenshot output.

Note: file:// URLs are blocked by Playwright MCP, so the HTTP server is required. Start it once and reuse it for all diff images in the session.

Reading the diff image:

• Red pixels = differences between baseline and current
• Yellow pixels = anti-aliasing differences (usually ignorable)
• Dimmed original = areas that match (shown at 30% opacity for context)
• Large red clusters = structural issues (missing elements, layout shifts, wrong sizing)
• Scattered red dots = sub-pixel rendering, font smoothing, or anti-aliasing (usually acceptable)
• Red bands at edges = dimension mismatch or padding/margin differences

Step 6: Interpret and act

Combine the quantitative JSON data with the visual diff inspection:

diffPercent	Interpretation	Action
< 1%	Visually identical	Minor sub-pixel differences only. No action needed.
1–5%	Close match	Inspect diff image for spacing, font weight, or border differences. Small CSS tweaks likely needed.
5–15%	Notable differences	Layout shifts, color mismatches, or missing elements. Review the red clusters in the diff to identify which sections need work.
> 15%	Significant gap	Major structural or styling differences. Focus on the largest red regions first — these indicate the biggest layout discrepancies.

When interpreting the diff image, note:

• Where on the page are the red clusters? (top, middle, bottom, left, right)
• Are they associated with specific elements? (navigation, hero, cards, footer)
• Do the clusters suggest spacing issues, color differences, or missing content?

Step 7: Fix and re-diff

1. Make CSS/HTML corrections based on the diff analysis
2. Repeat Steps 3–6 (only re-screenshot the current URL)
3. Continue until diffPercent reaches an acceptable level

Typical targets:

• Exact replication: < 2%
• Reasonable match: < 5%
• Structural match (different content): < 15%

---

Full-page screenshots

By default, screenshots capture only the visible viewport. For long pages, use the fullPage option:

mcp_playwright_browser_take_screenshot → {
  type: "png",
  fullPage: true,
  filename: "temp/vdiff-baseline-desktop-full.png"
}

Warning: Full-page screenshots produce large images and slower diffs. Only use when comparing entire page layouts.

---

Storybook-specific tips

When screenshotting Storybook stories:

1. Use the iframe URL for cleaner screenshots (no Storybook chrome):

   http://localhost:6006/iframe.html?id=components-card--default&viewMode=story
   

2. Wait for render — Storybook stories may take a moment to hydrate:

   mcp_playwright_browser_wait_for → { time: 3 }
   

3. Element screenshots work well for isolating the story content from any Storybook padding.

---

Threshold tuning

The --threshold flag (0 to 1) controls pixelmatch sensitivity:

Value	Sensitivity	Use case
0.05	Very strict	Exact pixel matching, catches everything
0.1	Default	Good balance, ignores most anti-aliasing
0.2	Lenient	Tolerates font rendering differences across systems
0.3	Very lenient	Only catches major color/layout differences

node scripts/visual-diff.mjs \
  --baseline temp/vdiff-baseline-desktop.png \
  --current temp/vdiff-current-desktop.png \
  --output temp/vdiff-diff-desktop.png \
  --threshold 0.2

---

Example: Comparing a hero section

# 1. Screenshot baseline (production site hero)
mcp_playwright_browser_navigate → { url: "https://www.bitovi.com/" }
mcp_playwright_browser_resize → { width: 1280, height: 720 }
mcp_playwright_browser_wait_for → { time: 2 }
mcp_playwright_browser_snapshot → {}
# Find ref for hero section, e.g. ref="hero-[1]"
mcp_playwright_browser_take_screenshot → {
  type: "png",
  element: "hero section",
  ref: "hero-[1]",
  filename: "temp/vdiff-baseline-hero-desktop.png"
}

# 2. Screenshot current (Storybook story)
mcp_playwright_browser_navigate → { url: "http://localhost:6006/iframe.html?id=components-hero--default" }
mcp_playwright_browser_resize → { width: 1280, height: 720 }
mcp_playwright_browser_wait_for → { time: 3 }
mcp_playwright_browser_take_screenshot → {
  type: "png",
  filename: "temp/vdiff-current-hero-desktop.png"
}

# 3. Run diff (in terminal)
node scripts/visual-diff.mjs \
  --baseline temp/vdiff-baseline-hero-desktop.png \
  --current temp/vdiff-current-hero-desktop.png \
  --output temp/vdiff-diff-hero-desktop.png

# 4. Start temp server (once per session, in background terminal)
npx -y http-server temp/ -p 8787 --cors -c-1 &

# 5. View diff image via Playwright
mcp_playwright_browser_navigate → { url: "http://localhost:8787/vdiff-diff-hero-desktop.png" }
mcp_playwright_browser_take_screenshot → {}   # triggers vision

# 6. Interpret: read JSON output + visually inspect diff
# 7. Fix styles, re-screenshot current, re-diff

---

Cleanup

All output files go to temp/ which is gitignored. To clean up after a session:

rm -f temp/vdiff-*.png