Genui — Design skill for Claude Code


A properly prompted modern LLM (Gemini 3) can robustly generate a full custom interactive web page for almost any prompt. Humans prefer it over markdown 82.8% of the time.

— summarized from Leviathan et al., [*Generative UI: LLMs are Effective UI Generators*](https://arxiv.org/pdf/2604.09577), Google Research ([blog](https://research.google/blog/generative-ui-a-rich-custom-visual-interactive-user-experience-for-any-prompt/), [examples](https://generativeui.github.io))

**genui is Claude Code without the chat window: a coding agent whose only output is interfaces it generates on the fly.** Ask about your code and, instead of a wall of markdown, you get a custom panel built for that answer: an architecture diagram, a diff to approve, a table of test results, buttons for next steps. You can click and comment on any part of it. It runs locally in your browser on your existing Claude Code login; run `genui` in a project to start.

It stays fast through orchestration: a strong model (Opus or Fable) does the thinking and writes a short brief, and pre-warmed fast models (Haiku) stream each brief into a panel. Panels start appearing in **~0.5–3 seconds instead of 1–2 minutes**. [Why this works →](#why-genui)


*Real-time capture, not sped up. Full video: [docs/demo.webm](docs/demo.webm).*

Contents

• Why genui
• How it works
• Install
• CLI
• Sessions
• License

Why genui

The paper shows people strongly prefer generated interfaces, but measured that *"when ignoring generation speed."* Its implementation has one model write each full page. genui keeps the paper's prompting principles (Appendix A.5) and post-processing ideas (A.6), and changes the architecture to remove both costs:

Paper's limitation	genui's answer
1–2 minutes per page (streaming only halves the perceived wait)	A strong planner writ