Forty-six files hide a labor-intensive cloning workflow
Our measured checkout had 46 files, about 148 lines of source, and occupied 1.4 MB. That sounds almost too small to clone a production site because the source code is not the main product. The repository is a clean Next.js scaffold plus a long instruction set that tells an AI coding agent how to inspect a target, record its findings, divide the build, and check the result.
The /clone-website command accepts a live URL and produces a front-end reconstruction. It does not recover the target's private source code. The skill directs the agent to extract computed CSS, real text, images, video, SVGs, responsive layouts, and interactive states from the rendered page. The output uses Next.js 16, React 19, Tailwind CSS 4, and shadcn/ui.
Three viewport widths make browser automation mandatory
The skill requires inspection at 1,440 px, 768 px, and 390 px. Its full cloning procedure also tells the agent to scroll, click, hover, and wait for timed behavior before building. A static page fetch cannot reveal a sticky header threshold, a hover transition, or whether a changing panel responds to scrolling rather than clicks.
Preflight stops when Chrome, Playwright, Browserbase, Puppeteer, or another browser automation tool is unavailable. That makes the README's one-command pitch conditional: the command starts a supervised process, not a deterministic converter. A model still has to notice every state, use the browser correctly, and preserve the target's route behavior. The saved screenshots and behavior notes give a developer something concrete to audit when it misses.
What happened when we ran it
Our sandbox installed commit 0fc4dca in 33 seconds with npm. The install added 889 packages and occupied 616 MB, a large dependency footprint for a 1.4 MB checkout. The run used a fresh unprivileged Debian container with 3 CPUs, 8 GB of RAM, Node 22, and no secrets. Installation completed successfully.
The production build passed in 25 seconds. Npm audit reported 0 known vulnerabilities: 0 critical, 0 high, 0 moderate, and 0 low. Those results apply to commit 0fc4dca, which pins Next.js 16.3.5 after a dependency security fix. We did not run an agent against a live target, so the build says nothing about cloning accuracy or browser behavior.
No test script or target existed, and our repository scan found no tests directory. The harness skipped tests rather than passing them. The project did have 1 CI workflow, a Dockerfile, and a Compose file. CI runs installation, linting, type checking, and the production build, but it does not exercise /clone-website against a reference page or score a visual diff.
A 150-line spec ceiling keeps builder jobs bounded
The skill tells the coordinating agent to split a component when its specification exceeds about 150 lines. Before dispatch, every builder receives exact computed styles, text, asset paths, interaction rules, and responsive behavior. Builders work in separate Git worktrees, run TypeScript checks, and return pieces for assembly. The coordinator then builds after merges and performs desktop plus mobile comparisons against the original.
That procedure is the template's best feature. It replaces a vague request such as "copy this page" with inspectable files and bounded tasks. It also assumes the agent can delegate work and manage worktrees without damaging existing routes. If your chosen client lacks those abilities, the process becomes more manual, and the README does not provide measured fidelity results for any of the 4 supported agents.
Five product layers stay outside the default clone
Five default exclusions matter: a real backend, authentication, real-time features, SEO work, and an accessibility audit. Mock data can make a copied dashboard look convincing while its forms, permissions, search, checkout, and data lifecycle remain absent. Treat the generated project as a front-end reconstruction that still needs product engineering, not as a recovered application.
The 25-second build checks the untouched starter, not the site you later generate. Every cloned route needs its own keyboard review, responsive checks, loading and error states, and production security work. The skill does require visual QA and build verification after assembly, but visual similarity cannot prove that a form is accessible or that an account boundary is safe.
Three prohibited uses put permission ahead of fidelity
The README names 3 prohibited uses: phishing or impersonation, passing off another design as your own, and reproduction that violates site terms. The published usage limits matter because the workflow intentionally downloads real text and assets. Ownership is a setup question, not a disclaimer to consider after the clone looks good.
A sensible target is a site you own, a migration you are authorized to perform, or a page used privately for study. Logos, photographs, copy, fonts, and videos may each carry separate rights. Even with an MIT-licensed template, those downloaded materials do not inherit the template's license. Review the target's terms and your permissions before the agent saves the first asset.
v0.5.1 and a September 27 push show current maintenance
Release v0.5.1 shipped on September 27, 2026, one week after our measured commit. Its notes say Next.js was pinned at 16.3.5 and the README was updated to recommend Opus 5.5. GitHub showed 35,398 stars and 1 open item on September 28; that item was a pull request rather than a reported issue. The repository was last pushed September 27.
AI Website Cloner Template is worth using when the inspection record matters as much as the first screenshot. Its 0 audit findings and passing 25-second build make the starter easy to trust as a base. The generated site still deserves a fresh test plan, permission review, and manual comparison, because the repository has no automated result showing that one live URL becomes a faithful, functional application.