design2code-rl
SourceAn RL environment that grades coding agents on rebuilding multi-page websites from screenshots
Claude Code with Opus 4.7 rebuilt multi-page websites from screenshots with a mean reward of 0.80 over 100 trials. The oracle, which feeds the original site back to the grader, scored 0.98.
Why build it
Screenshot-to-code benchmarks such as Design2Code and Web2Code test one page. Real sites share a nav, a footer and a brand across pages, and drift between pages is a failure a single page cannot show. I also wanted a reward that could train a model as well as rank one. Pass or fail gives no gradient, so the grader returns a score between 0 and 1.
How it works
The pipeline generates 50 sites of five or more pages. Claude Sonnet writes each concept (a category, a brand and a design language), Claude Opus writes the HTML and CSS with shared regions first, and a screener rejects broken sites once Playwright renders them.
Each site is a Harbor task run on Modal. The grader re-renders the agent's pages and scores:
- Visual fidelity (70%): three Opus 4.7 judge calls compare desktop and mobile screenshots of each page
- Text (15%): headings, nav labels and body copy
- Code quality (7.5%): valid, semantic, responsive and accessible markup
- Consistency (7.5%): brand colours across pages
The judge sees only screenshots, so any markup that looks right scores well. The eval ran Claude Code 10 times on each of 10 varied sites.

Findings
The site decided the score far more than the attempt did. Within a site the standard deviation stayed under 0.05, while site means ran from 0.91 for a minimal space nonprofit to 0.67 for a retro storybook education site.
Retro was the hardest design language, about 0.18 below minimal, because the agent dropped decorative type, background patterns and ornaments. About 60% of the judge's lost points came from photos and illustrations the agent cannot draw. Accessibility scored 0.44, mostly for missing alt text and heading order.
Details are in the visual report, failure patterns and grading calibration.