cowboycodr/sitegeist

Sitegeist is a visual benchmark that asks a simple question: when leading coding agents get the same set of website briefs, how often do they land on the same design choices? The project contains 100 neutral website briefs and 400 finished, self-contained websites, generated by four different model collections named 5.6 Sol, Opus 4.8, Grok 4.5, and Fable 5. A gallery lets you browse every result, switch between models, and compare up to three generations side by side. What sets this benchmark apart is how each site is produced. Every generation runs in a fresh, isolated workspace. The model doing the work only sees a neutral work order, one private brief, and the rules for what counts as a valid submission. It cannot see the Sitegeist gallery, the project's git history, another model's result, earlier sites from its own batch, or any screenshots or feedback from previous attempts. Accepted sites must include their original source code, build to a plain static site with no external dependencies, and come with a self-contained poster image for the gallery. Submissions get rejected if they load resources from the internet, depend on a live network connection while running, use symbolic links, are malformed, throw errors in the browser console, or overflow the screen on mobile. The goal is not to crown a winner on taste, but to preserve the outputs and the strict boundaries under which they were made, so anyone can inspect visual quality, repeated patterns, and exact copies for themselves. To run the gallery on your own machine you need Node.js and pnpm, then install dependencies and start the dev server, after which it opens at a local address. The gallery itself is built with SvelteKit and deployed on Cloudflare Workers. The project also ships deterministic audit scripts that check for exact duplicate code or builds, duplication across different models, how much content each collection contains, and whether a finished collection is internally consistent. The README is upfront about the benchmark's limits. It measures one narrow website generation task, not overall model capability. Its duplicate checks catch identical source or output, not looser stylistic similarity. The gallery poster images are just representative screenshots, not the actual evidence, since the real sites stay viewable in full screen. It is released under the MIT License.