website-rebuild-skill rebuilds sites and proves the copy matches pixel for pixel

boyang-hu's website-rebuild-skill is an Agent Skill that gives a URL to an AI agent, reconstructs the site's implementation line by line from the minified code, and grades the copy against the original with a set of automated acceptance checks. It was released on 11 August under MIT and is at 941 stars as of 6 September.

The skill implements the Agent Skills open standard and the README documents runs completed end-to-end by both Claude Code and Codex, including one target where Codex produced 166 out of 166 identical response bytes on the Hashgraph VC site. The same skill directory runs the same pipeline in whichever runtime it runs in.

Five verification checks

The pixel diff is one check of five. The full set of automated comparison layers is console output, network traffic, DOM tree, geometry, and pixel-by-pixel image comparison. Every difference is either fixed or logged; the pipeline refuses to hide one without reporting it. The verified showcases list nine sites, including a 1.25 MB custom WebGL engine (Lusion) with 156 shaders, and the Apple AirPods Pro landing page rebuilt from 565 webpack modules with nine scroll checkpoints pixel-identical. Where the rebuild cannot reach meanAbsDiff 0.00, the README states the exact same-side noise band the rebuild fell into.

There is one rule in the code that matters more than the score. The rebuild copies bugs and odd states verbatim, and never "fixes" them. An unloaded lazy-image placeholder strip on the Raycast keyboard page is identical on both sides in the acceptance shot. The point is behaviour parity, and a fix applied without notice breaks the proof.

Publishing is off by default

The author is direct about the ethics. Output defaults to private, noindex, and not deployed. The skill only collects evidence and lists options with their risks. Any decision that touches "publish, deploy, or redistribute" is left to the operator. The MIT licence covers the skill's own code; the copyright of any site it reconstructs belongs to that site's owner.

The reasons an agency, platform team, or design system group would trial the skill are competitive analysis, recovering a site that no longer runs from the Wayback Machine (the README lists five recovery attempts, four revived, one lost with the failure mode written up), and turning a large framework build back into a readable project. On a 449,000-line Nuxt or Vite site, the skill split the output into 2,043 semantically named parts across scene, camera, wave and other folders, then rebuilt it byte-for-byte.

The pipeline uses 75 Node scripts, installs no npm packages until the final stage, and asks the operator up front which of three completion levels to target. A mirror archive, an engineered rebuild, or the source-form port; the levels build on each other, so choosing the cheaper level today does not block an upgrade tomorrow.

Source

boyang-hu/website-rebuild-skill on GitHub. Star count and repository dates from the GitHub REST API, 2026-09-06. Verification numbers and copyright policy from the English README.