Reproduced features
Known limitations
Verbatim prompts from the experiment
Loading…
We gave coding agents real websites to rebuild. Here are the original prompts, visual evidence, source code, and the parts they couldn't reproduce.
Documented tests · No invented success scores


Select an experiment to inspect desktop/mobile screenshots, original prompts, working features, and known limitations.
Loading…
Codex and Claude Code independently attempted the same iPhone 18 Pro page using the website-cloning skill. Both documented their work and limitations.

3D/AR and exact scroll animation timing are not reproduced. Film playback needs Apple online.

Videos play when in view rather than following scroll frame-by-frame. 3D/AR isn't reproduced.
Codex has labeled paired source/clone screenshots. Claude's raw captures do not reliably label each browser tab. No numeric accuracy score is claimed without measurement.
Getting a successful build is not the same as accurately reproducing a website. We document both the results and the gaps.
Study original pages on desktop and mobile, including menus, assets and scroll states.
Produce editable code, not a screenshot embedded as a finished website.
Examine matched screenshots and exercise interaction states when evidence is available.
Document missing 3D, approximation of animations, external links and server-side features.
Browse the complete files and original prompts in the repository.