I've been exploring how to describe UI layouts to LLMs efficiently.

The problem: When you ask an AI to generate or modify UI, how do you describe the current state? - Natural language ("header on top, form below") is ambiguous - ASCII art breaks when edited (alignment issues) - HTML is precise but verbose

I ran some measurements. For a simple login form: - Natural language: 102 tokens - ASCII art: 84 tokens - HTML: 330 tokens

I experimented with a grid-based text format using Excel-like cell references:

  grid: 4x3
  A1..D1: { type: txt, value: "Login" }
  A2..D2: { type: input, label: "Email" }
  D3: { type: btn, value: "Submit" }

This came out to 120 tokens – less than HTML, more precise than natural language.

Built a CLI to render it to SVG/PNG: npx ktr input.kui -o output.png

Curious what approaches others have tried for this problem. Is there something I'm missing that already solves this well?

Code: https://github.com/enlinks-llc/katsuragi