How do you keep AI-built UI consistent?

When I vibe code with Claude Code or Cursor, the agent tends to invent a new button, card and table on every screen. The app works, but each page looks slightly different, and sometimes the agent builds a placeholder UI and calls it done.

A few things I've tried:

• Putting design rules in the prompt or an AGENTS.md file. It helps, but the agent drifts after a few screens

• Giving it a component library to use. More consistent, but only if the agent can actually find and read the components

• Fixing it by hand at the end. Works, but slow

What's working for you? Do you give your agent a design system, a component library, a reference screenshot, or something else? And which AI tool keeps UI most consistent in your experience?