The Age of Building Screens with One Prompt Line, and How to Make a 'Daangn-like Screen'

AI UI tools can create plausible-looking screens with just one line of prompt, but they couldn't match the component, token, layout, and copy rules of the SEED design system. If generic Button or Modal came out instead of SEED components like ActionButton, AppBar, and BottomSheet, they couldn't be directly inserted into the Daangn app, and Lovable, v0, Bolt, and Figma Make were also tied to their own components or Figma canvas.

Kraft was designed not as a tool that draws screens to solve this problem, but as a tool that automates the decisions required by the design system.

• Admin: It provided a prompt → preview → deploy → webview check flow by simply opening a URL with no installation needed, but it was a generic structure that used the same agent for all teams.
• CLI: By having it read local projects and existing code, it enabled team-specific context and session separation, and bundled the editor frontend and agent backend together so it could run with a single npx line.
• Agent: Through the Clarify → Memory Read → Cases → Generate → Verify → Memory Write loop, it concretized ambiguous requirements, verified SEED rule compliance and required state handling, and then accumulated memory.

The implementation mixed a Harness combining two modes, Plan / Orchestra, with delegation to Claude Code (Claude Agent SDK). The design is stored separately from the code as DesignSpec JSON, and the actual coding spawns the local claude binary to handle file modification and building. In the end, Kraft evolved not into a 'one-time screen drawing tool' but into a device that makes the design system's judgment repeatable.