Codex Takes the Design Wheel: A Hands-On UI Reskin of nvm-desktop
This workflow lets an AI handle the bulk of UI implementation on an existing app while the human reviews rendered pages and gives targeted visual feedback. The three-layer test setup turns the AI into a self-verifying contributor, cutting down the manual QA loop and making AI-generated design changes more trustworthy.
Codex, running GPT-6.1 Sol in Light mode, got a one-line brief: simple, modern, premium. No mockups, no spacing or color specs. The developer configured three skills (shadcn for components, UI UX Pro Max for design reference, and a custom comment skill), then let Codex produce a first pass. From there the loop was straightforward: Codex edits, the developer reviews the code, opens the actual page, and points out specific problems like crowding or visual imbalance.
The surprise was that Codex understood nvm-desktop's existing theme-switching system and tuned the design across themes, so changing themes didn't break the look. The author also added three layers of automated testing (Vitest in-memory, Playwright real-browser, and WebdriverIO E2E) so Codex could write and run its own test scripts after each change. The human still manually inspects pages for visual comfort and interaction smoothness.
The human's role shifts from writing front-end code to acting as a design critic after seeing concrete output, which makes feedback much more specific than abstract style guidance.
Automated tests matter more when an AI is doing the implementation because they let the AI close its own loop, leaving the developer to focus on visual and UX judgment that tests can't capture.
The fact that Codex handled theme switching without explicit instructions suggests LLMs can infer and respect existing architectural constraints when given enough context, even in design work.