Designing with AI works best as a loop: you describe, the AI drafts, you judge and refine, and you repeat. The AI is fast at producing options. You are responsible for choosing the right one.
Step 1: write a brief. Describe the user, the task, the content and any constraints. "A settings screen for a budgeting app, for people on phones, with sections for account, notifications and privacy, using our existing components" will give a far better result than "make a settings page".
Step 2: generate drafts. Prompt-to-UI tools such as Google Stitch or UXPin's AI assistant turn a description or a sketch into screens. Ask for several variants and compare them, rather than accepting the first one.
Step 3: refine in a design tool. Bring the best draft into Figma, Penpot or another editor. Fix hierarchy, spacing and copy, and design the states the AI usually skips: loading, empty, error and disabled.
Step 4: connect to code. AI coding tools such as Claude Code and Cursor can build the interface. Results are much better when they can read your design through an MCP server. Figma and Penpot both have one. For interactive and animated parts, ExodeUI, which we make, has an MCP server that lets the AI create and edit animations with states and spring physics from plain-language prompts, stored in an open .exode file your app plays.
Step 5: review before you ship. Check keyboard and screen reader use, contrast, reduced motion, small and large screens, and long or translated text. Treat AI output like work from a fast but new teammate: useful, but always reviewed.