rubenhassid1/remake

Remake is a Claude Code skill/plugin and command-line engine for rebuilding a public website as a copy you own, then replacing its text, images, logos and fonts with your own. It is aimed at developers and AI-agent users who want to learn how a site is built, migrate their own site, or start a new one from a look they like.

What it does

One prompt in Claude Code (/goal Remake <url> ...) drives the whole process: capture, rebuild, score, repair, ship. The output includes a local site/ folder, a private GitHub repo, a Vercel deployment behind a passcode and hidden from search engines, a side-by-side comparison page at /_remake/, a swap board at /_remake/swap/ listing every remaining original element, and a report ending in REMAKE COMPLETE or the remaining gaps (README).

How it works

The engine visits the page once at desktop size and once at phone size, records HTML, CSS, scripts, fonts, images and the API responses the page made, then replays from that recording with rewritten addresses and trackers stripped. Its distinguishing step is deterministic capture: it pauses the page clock, seeds random-number calls, holds video on the first frame and steps time forward itself, so two loads produce comparable pixels. The copy is then scored at up to 40 screens per page plus 24 motion frames, requiring 98% of stable pixels, 95% per motion frame and page heights within 4 px. Claude reads diff images and repairs until these gates pass (README, how it works).

Because the original markup, styles and scripts run in the copy, animated and WebGL-heavy sites keep their effects; this also means the result inherits the original's behaviour. The swap step fills slots from remake/content.json; the authorship claim is that original photos are measured only for shape, colour and light, and replacement visuals come from generation or public-domain stock.

Ways to run it and requirements

There are three paths: prompt-only (Claude reads the repo and installs the engine into ~/.remake), a Claude Code plugin (/plugin marketplace add ...), or the CLI (npm install -g from GitHub). All paths need Claude Code, Node 20+, git, GitHub CLI, and free GitHub and Vercel accounts; the first run also downloads around 150 MB of Chromium for Playwright. If the global install fails on permissions, the documented fallback is running the engine directly with Node. Setup details, including remake doctor --fix, are in the README; the agent-facing skill is in SKILL.md.

Limits and conditions

Remake deliberately stops at password or card fields, so sign-in, checkout and banking pages are not copied. It does not evade bot walls, use proxies or solve CAPTCHAs, and rebuilding from the Internet Archive is an explicit choice rather than automatic. Pages behind logins, sites that block automated visits, live-changing data, WebSocket apps and checkout flows are documented as unsupported. Backend behaviour is not copied; API responses seen during capture are replayed, and forms are held back. Publishing is constrained by design: remakes start private, and the author says to keep them private until the swap board reads 100%. The original's content and code remain the owner's property.

Timing varies by site: a simple marketing page may pass on the first score, while an animated agency site may take hours in the repair loop. The project is MIT-licensed (repository).