Perfect-Web-Clone

Perfect Web Clone

Pixel-perfect clones of any webpage. This repo is the measured core. The product is the skill: an agent harness that turns one clone into a full-page, scored replica.

English | 中文

A screenshot is not a clone. The skill drives the agent through capture → section plan → clean React → build → measured repair, and it will not stop after the first viewport or call a similar-looking inert widget “done.”

Say this to your coding agent:

The skill is the product

skill/SKILL.md is the harness. Same file is published as perfect-web-clone-skill. One request authorizes the whole run Hands and eyes are deterministic (pwc) — they never call a model The agent authors code and repairs the worst measured section Done means ready_for_user_review or failed_with_residuals, with evidence Pixel-perfect is a gate table, not a prompt Install

Python 3.10+ and Node 20+.

Load perfect-web-clone-skill into Claude Code, Codex, or any coding agent that can run pwc. How a clone run works

1. Capture the live URL (integrity-checked, assets localized) 2. Plan the real page sections 3. Author clean React for every section 4. Assemble and build 5. Score fingerprints, weight, and visual match 6. Repair the worst section, rebuild, re-measure 7. Hand you a local preview: ready_for_user_review or failed_with_residuals

If the page is WebGL, runtime canvas, or scroll-choreographed, that ceiling is reported up front. Content still clones; the runtime-drawn look may not. CLI

Every command prints JSON. ok is the only pass/fail that matters. Layout