bitovi/ai-enablement-prompts/pixel-perfect

Skill: Pixel-Perfect Orchestration

Drive a page or component to pixel-perfect fidelity with a baseline by orchestrating visual-diff (screenshot comparison) and computed-styles (CSS property extraction) in a converging loop.

When to Use

• Replicating a production page's look in the local codebase
• Matching a Storybook component to a live reference
• Any task where the user says "make it match" or "pixel perfect"

When NOT to Use

• Building new designs from scratch (no baseline exists)
• Content-only changes with no visual target
• Figma-based design work (use the Figma MCP skill instead)

Prerequisites

1. All prerequisites from the visual-diff and computed-styles skills are met
2. The baseline URL and current URL are known
3. The dev server or Storybook is running

Design System

All fixes must use the Bitovi design token system defined in src/styles/bitovi-theme.css. Prefer token-based Tailwind classes over arbitrary values. Only use [arbitrary] syntax when no matching token exists in the theme.

Workflow Overview

┌─────────────────────────────────────┐
│  For each breakpoint:               │
│  desktop → tablet → mobile          │
│                                     │
│  1. VISUAL DIFF                     │
│     Screenshot both pages           │
│     Run pixel diff                  │
│     Identify problem regions        │
│     diffPercent < 2%? ──── YES ──── next breakpoint (or DONE if all 3 pass)
│              │                      │
│              NO                     │
│              ▼                      │
│  2. COMPUTED STYLES                 │
│     Extract CSS from both pages     │
│     Build property diff table       │
│     Fix mismatches with Tailwind    │
│     Verify fixes took effect        │
│              │                      │
│              ▼                      │
│  3. RE-DIFF                         │
│     Re-screenshot current page      │
│     Run pixel diff again            │
│     diffPercent < 2%? ──── YES ──── next breakpoint
│              │                      │
│              NO                     │
│              ▼                      │
│     Go to step 2 with new regions   │
└─────────────────────────────────────┘

Detailed Steps

Step 1: Initial Visual Diff

Load the visual-diff skill and run the full workflow:

1. Screenshot the baseline at desktop breakpoint (1280x720)
2. Screenshot the current page at the same breakpoint
3. Run node scripts/visual-diff.mjs to produce the diff image and JSON stats
4. View the diff image through the HTTP server + Playwright vision
5. Record the diffPercent and identify which regions (top, middle, bottom) have red clusters

If `diffPercent` < 2%, the pages are effectively matching. Report success and stop.

If `diffPercent` >= 2%, proceed to Step 2.

Interpreting the Diff Image

Map red clusters to specific DOM regions:

Step 2: Computed Style Extraction

Load the computed-styles skill. Based on the regions identified in Step 1:

1. Plan extraction targets: Map each red region to specific DOM elements and CSS properties
2. Extract from baseline: Use mcp_playwright_browser_evaluate on the baseline URL
3. Extract from current: Use mcp_playwright_browser_evaluate on the current URL (remember the Storybook iframe pattern if applicable)
4. Build diff table: Compare every property, flag mismatches
5. Fix mismatches: Apply Tailwind classes — always resolve computed values to design tokens from src/styles/bitovi-theme.css first (see the Design System section above). Only use [arbitrary] values when no token matches.
6. Verify each fix: Re-extract from current to confirm the computed value now matches

Prioritization

Fix differences in this order (highest visual impact first):

1. Layout: width, height, maxWidth, padding, margin, display, flexDirection
2. Typography: fontSize, fontWeight, fontFamily, lineHeight, letterSpacing
3. Colors: color, backgroundColor, borderColor
4. Decoration: borderRadius, boxShadow, opacity
5. Fine-tuning: textTransform, gap, overflow

Step 3: Re-Diff

After fixing a batch of computed style differences:

1. Re-screenshot only the current page (baseline hasn't changed)
2. Re-run the pixel diff
3. View the new diff image

Evaluate progress:

Step 4: Convergence

Continue the loop (Steps 2-3) until diffPercent < 2% or the remaining differences are confirmed to be not worth further pursuit.

Stop and flag for manual review — do NOT keep iterating on these:

• Sub-pixel font rendering (different OS/browser/font-smoothing)
• Anti-aliasing artifacts (icon edges, curved borders)
• Pixel-level positioning offsets (elements within 1–3px of correct position)
• Shadow or gradient rendering variance
• Third-party embedded content (reCAPTCHA, chat widgets, analytics overlays)
• Dynamic content (timestamps, user-specific data, live chat, cookie banners)
• Differences that require non-Tailwind workarounds or browser-specific hacks

When stopping for one of these reasons, clearly state:

1. What the remaining diffPercent is
2. Which specific element(s) are causing it
3. Why it is not worth automating further
4. What the user could do manually to address it

Using Subagents

For complex pages with many regions to fix, use subagents to parallelize investigation:

Investigation Subagent

Use the Explore agent to research element structures before extracting styles:

"Explore the DOM structure of [baseline URL]. Find all elements in the 
[region] section. Report element tags, class names, and nesting structure. 
Thoroughness: medium."

Extraction Pattern

When extracting styles, do both pages in sequence (not parallel, since Playwright MCP shares one browser):

1. Navigate to baseline → extract → store results
2. Navigate to current → extract → store results
3. Compare in conversation context

Breakpoint Strategy

Always run the pixel-perfect loop at all three breakpoints:

Workflow:

1. Run the full visual-diff + computed-styles loop at desktop until diffPercent < 2%
2. Then run at tablet — resize viewport, re-screenshot both sides, diff, fix, re-diff
3. Then run at mobile — same process

Fix desktop first so you're not chasing responsive issues against a broken base layout. Each breakpoint has its own screenshot files (e.g. temp/vdiff-baseline-tablet.png, temp/vdiff-current-tablet.png, temp/vdiff-diff-tablet.png).

Component-scoped comparison at smaller breakpoints: Re-measure the component height at each breakpoint (it will change) and resize the viewport accordingly before screenshotting.

Session Notes

Keep a running log of what was fixed and verified. This helps avoid re-checking properties and provides a summary at the end.

Use session memory (/memories/session/) to store:

• The diff table with baseline vs current values
• Which properties have been fixed and verified
• Current diffPercent after each round

Success Criteria

The default target is < 2%. Stop and report as soon as this threshold is met — do not keep iterating to squeeze from 1% to 0% unless the user explicitly asks. Flag remaining issues that fall into the "hard to align" bucket (font rendering, sub-pixel spacing, third-party overlays) and let the user decide whether to fix them manually.

Example Session

User: "Make the nav component match production"
 
Agent:
1. Load pixel-perfect skill
2. Load visual-diff skill → screenshot production nav + Storybook nav
3. Run diff → diffPercent: 12.3%
4. View diff → red clusters at: nav height, logo area, nav item spacing, CTA button
5. Load computed-styles skill
6. Extract nav properties from production:
   nav.height = 80px, nav.padding = 0px 128px, logo.width = 108px ...
7. Extract same from Storybook:
   nav.height = 60px, nav.padding = 0px 32px, logo.width = 80px ...
8. Build diff table, fix mismatches with Tailwind — resolve to design tokens from `src/styles/bitovi-theme.css` (e.g. `bg-brand-orange`, `text-teal-500`, `font-semibold`, `p-md`)
9. Verify fixes with re-extraction
10. Re-screenshot + re-diff → diffPercent: 3.1%
11. View diff → remaining red at CTA button border-radius
12. Extract CTA styles, fix borderRadius
13. Re-diff → diffPercent: 0.8%
14. Done! Report final comparison table.

Cleanup

After the session:

rm -f temp/vdiff-*.png

All computed style data lives in conversation context and session memory only.