My setup changed a bit. I now work with Claude Code in VS Code, and probably my newest discovery is Superpowers. I mostly use the writing-plans skill and then let it develop subagent-driven in separate worktrees. That way three to four tickets get coded in parallel. It uses a lot of tokens, but I find the output really promising.

To keep the UI consistent, I rely on Storybook and run visual regression tests. On top of that, the CLAUDE.md states that every new component must have a story. This works quite well for me and helps me avoid unintended UI changes.

I first created the wireframes in Claude Design, based on the specs I had at the start of the project. Once that was sorted, I had a small, simple design system created — it lives in Storybook too, along with the design tokens. New components are now built directly in Claude Code using the tokens that are available.