Project hub
Claude Design vs Open Design: Website Build
A live comparison of Claude Design (Opus) and Open Design (Codex) while building the Jason Vs The Noise website: taste, speed, implementation quality and what each agent is actually best at.
This is an operating journal, not a benchmark paper. The verdict is based on this website build, not a universal model ranking.
This project tracks the build of the Jason Vs The Noise website as a real-world comparison between two different agent strengths:
- Claude Design (Opus) for taste, visual direction, hierarchy and avoiding generic AI-design sludge.
- Open Design (Codex) for implementation, Astro/MDX production build-out, routing, debugging and turning the design into a working site.
The question
Which agent actually wins when the job is not a benchmark, but a public website that has to look good, build cleanly and be easy for agents to keep publishing into?
The result so far
Winner for visual direction: Claude Design (Opus).
It produced the stronger design judgement: the warm editorial grid, confident type scale, premium card system and overall “proper website” feel. That was the breakthrough after the app-builder versions looked too much like dashboards wearing a fake marketing site costume.
Winner for production build-out: Open Design (Codex).
Codex turned the direction into the Astro site: content collections, project pages, diary routes, resource pages, active nav states, build fixes, internal-link checks and browser QA. Less glamorous, more useful. The spanners matter.
Current verdict
The winner is not one agent. The winner is the workflow:
Claude Design for taste. Open Design for build. Astro/MDX for the spine.
That combination got the site from rough idea to something Jason actually loved in the browser.
What this project will track
- Which design decisions came from Claude Design.
- Which implementation decisions came from Codex.
- Where the hand-off worked cleanly.
- Where agent output needed human taste or operator judgement.
- How fast new pages and journals can now be added by dropping MDX files.
The useful lesson
AI agents are not interchangeable magic interns. They have different postures.
Some are better at taste and direction. Some are better at getting the bloody thing wired up. The leverage comes from using each one where it is strongest, not pretending one model should do every job equally well.