🧬 clone-team

Clone any website with a coordinated team of AI agents — and get the architecture docs to rebuild it.

A Claude Code skill that orchestrates a Manager, Frontend Developer, Backend Architect, and Tester to produce a pixel-perfect UI clone and a reverse-engineered ARCHITECTURE.md — in the stack you choose.

🌐 Website, demos & docs → clone-team.varalix.com

If this saves you a week of pixel-pushing, drop a ⭐ — it genuinely helps.

Original (left) vs. clone-team build (right) — wembi.ai, driven in lockstep so both stay at the same scroll point. A hard target: Lenis smooth-scroll + scroll-driven 3D + reveal animations. Full case study →

---

Why clone-team?

Most "AI clone this site" tools hand you a rough, one-shot approximation and call it done. clone-team is different in two ways:

1. It can't skip the quality gate. The build/test loop is a deterministic program, not a model's good intentions. A Tester runs a full visual + behavioral regression every round, and a section isn't "done" until the Tester says so — because "done" is literally a loop condition in code (see the engine).
2. You get understanding, not just pixels. A Backend Architect documents the target's routes, API/network surface, inferred data model, auth/session flow, and end-to-end user journeys — so you can build your own product on top, not just stare at a copy.

✨ Two deliverables, every run

	What you get
🎨 An exact UI clone	Pixel-perfect, behavior-accurate, in your stack (React, Next.js, Vue, Svelte, Angular, plain HTML/CSS…) — independent of the source site's stack.
📐 ARCHITECTURE.md	Routes, observed API/network surface, inferred data model & entities, auth/session flow, state & navigation, and step-by-step user journeys.

⚙️ The engine: a dynamic Workflow

clone-team is built on Claude Code's new dynamic Workflow engine — and it's the whole reason the quality guarantee holds.

A dynamic Workflow is a deterministic JavaScript program that spawns and coordinates a fleet of subagents — where the loop count, fan-out width, and exit conditions are all computed at runtime from real data, not hardcoded. clone-team uses it so the process can't drift:

• The Tester gate is unskippable. Per page: extract → spec → build → full-regression-test → fix → re-test … runs as a while loop whose exit condition is the Tester's OK. No human, no model, and no "looks fine under deadline pressure" can bypass it.
• It adapts to the work. One-page site → one builder. Nine-page site → nine, running concurrently. The Workflow reads the page list and fans out accordingly.
• It sizes the agent fleet to your PC. Before launching, a capacity probe reads your machine's free memory and computes how many builder agents to run at once — many in parallel on a big box, a few at a time on a laptop — so it goes as fast as your hardware allows without ever OOM-crashing it.
• It feeds failure back as input. A Tester NG isn't just a fail; its structured issue list becomes the next Developer round's fix list.
• It's pausable & resumable — even across sessions and usage-limit cutoffs — because progress is reconciled against a durable state.json + on-disk artifacts.