I haven't tried this myself, but from speaking with peers who are deeper into coding with LLMs, they recommended to try out something like Playwright MCP.
The idea is that you'd use Playwright as a way to automatically give "visual" feedback to the agent, so that it can a) get more accurate with designs it might implement with your chosen frontend styling tech (tailwind, UI kits, etc) and b) be able to iterate on a design without you constantly feeding images into it.
