DOM Clone

Rebuild the DOM/CSS layer of a page as clean components whose every value is measured, not estimated. The difference between a clone that reads as "pixel-perfect" and one that reads as "AI slop" is almost entirely whether the spacing, type scale, and colors are the site's real computed values or an approximation. So the whole method is built to remove guessing.

Scope boundary — read this first. This track handles DOM/CSS only. Any <canvas> painting WebGL/WebGPU/2D pixels is invisible to getComputedStyle() and belongs to shader-extract. When you meet a GPU surface (the surface map tags them, or you see a <canvas> you can't style your way into), do NOT rebuild it — leave a placeholder mount (below) and move on. Rebuilding a shader as HTML is the classic failure that makes a clone look empty.

Inputs

• Driven by clone-site: you get an output dir containing surface-map.json and ROUTING.md. Build only the sections routed to the DOM track; leave mounts for GPU surfaces.
• Standalone: you get a URL. Do a quick recon first (screenshots + the interaction sweep below). If you spot GPU <canvas> surfaces, tell the user those need the shader-extract track and build the DOM around them.

Requires a browser MCP (Chrome/Playwright/Puppeteer). Keep large extraction artifacts in the output dir, not the conversation.

The method: extract → spec → build → verify

1. Foundation first (sequential, do it yourself)

Nothing can be built until the shared ground exists, because every later component depends on it:

• Design tokens — run ../clone-site/scripts/tokens-probe.js (tokensProbe()) at 1440/768/390 if the coordinator hasn't already left tokens-*.json in the output dir. Its frequency-ranked palette, type scale, spacing, radii, shadows, and :root custom props ARE the tokens — write them as CSS variables / Tailwind theme, keeping the site's own --names when it has them. Colors and fonts wrong here poison every component.
• Motion params — likewise take motion.json (../clone-site/scripts/ motion-probe.js) as the source for every GSAP/ScrollTrigger/Lenis/WAAPI value a spec needs. A spec that says scrub: 1.2, start: "top top", end: "+=200%" came from the runtime; one that says "scrubs on scroll" is a guess.
• Fonts — identify actual font-family on headings/body/labels; wire up the same families (self-hosted or Google) so metrics match.
• Global assets — favicons, logos, and site-wide images to the project. Resolve each through the asset ladder in references/asset-resolution.md: download the REAL file first; only fall back to reconstructing, generating (Higgsfield, for decorative assets — never brand identity), or a placeholder, and label which tier was used. This keeps the clone honest about what's real.
• Global behaviors — smooth-scroll library (Lenis/Locomotive), scroll-snap, global keyframes. See references/interaction-fingerprints.md to detect and reproduce these; they change how sections must be built.

Verify the empty shell builds before going further.

2. Extract each section with exact values

For each DOM section top-to-bottom, use the browser to capture — don't hand- measure:

• A section screenshot (for the builder's reference and later QA).
• Computed styles for every meaningful element via getComputedStyle() — real px/color/weight values, not "looks like text-lg".
• Real content — verbatim text, alt text, and every image/<img>/background image in the section, including layered/overlay images (a hero is often a background + a foreground mockup + an icon; miss the overlay and it looks bare). Resolve each asset through references/asset-resolution.md (REAL → RECONSTRUCT → GENERATE → PLACEHOLDER), tagging any that couldn't be downloaded.
• All states, not just the default — for tabs/accordions click each and record per-state content; for scroll- or hover-triggered changes, capture the before AND after computed styles and the transition (duration/easing) plus the exact trigger. See references/interaction-fingerprints.md.

3. Write a spec file — the contract

Write one spec per section to <output>/specs/<section>.spec.md BEFORE building. This is the contract between measurement and construction: the builder works from the spec, not from its memory of a browser session, so it can't drift into guessing. A spec has: target file, screenshot path, interaction model (static / click / scroll / hover / time), exact computed styles per element, per-state content, assets with local paths, verbatim text, and responsive behavior at 1440 / 768 / 390. If a builder would have to invent any value, the spec isn't done yet.

For GPU surfaces in this section, the spec says only: placeholder mount — a positioned empty container with the id, size, z-index, and pointer-events from the surface map, for shader-extract to fill later. Nothing more.

4. Build — size agents to the model, not to dogma

Dispatch builders that each receive their spec inline (never "go read the spec"). On agent granularity: splitting a page into many tiny builders is a choice for speed and edit-isolation, not a fidelity requirement — a capable model holds a whole section's spec fine. So prefer one builder per section by default, and split further only when a section contains genuinely independent pieces (three distinct card variants, separate interactive widgets) that benefit from parallel isolation. Over-splitting just adds merge and coordination cost.

If builders run in parallel and might touch shared files, isolate them (git worktrees or separate output subtrees) and merge with the section topology in hand. Each builder verifies its own typecheck/build before finishing.

5. Verify with a real diff

After assembly, run an automated pixel diff (Playwright toHaveScreenshot()), original vs clone, at all three breakpoints — masking any GPU placeholder regions so their emptiness/animation doesn't dominate the result. Tune maxDiffPixelRatio rather than demanding zero diff (anti-aliasing causes false positives). For each real discrepancy, decide: was the spec value wrong (re-extract, fix spec) or did the builder ignore a correct spec (fix component)? Fix at the source, not with a compensating fudge.

Output

• site/ (or the coordinator's chosen substrate) with styled sections + mounts
• specs/*.spec.md — one per section, the audit trail
• Downloaded assets under the project's public dir
• QA diff results per breakpoint + list of placeholder mounts left for GPU track

What not to do

• Don't clone from a screenshot; extract the real DOM/CSS/content.
• Don't rebuild a <canvas> shader/3D effect as HTML — leave a mount.
• Don't approximate a computed value you could have measured.
• Don't freeze a <video> or animated SVG to a static image — re-embed / keep it.
• Don't build a click-based UI when the original is scroll-driven; settle the interaction model before building (it's a rewrite, not a tweak, to change later).