Hi Product Hunt 👋 I'm Divyansh from Avyukt Tech Labs. Most of the React UI I ship now is written by an agent, and most of what the agent writes isn't the UI. When I measured five common screens, 57% of the TSX tokens were class names, imports, prop types and closing tags. Terse is a small language for the part agents write most: what a screen shows. • One node per line. Styles are tokens like gap=md, so gap-[13px] can't even be written. • Data and handlers come in as props from normal TSX. State, effects and fetching stay where they are. • A Vite plugin compiles it at build time to plain shadcn/ui + Tailwind TSX. No runtime, nothing to eject. • Typos fail the build with file:line:col and "did you mean 'button'?", so the agent fixes it in one edit. On the five benchmark screens it's 755 tokens vs 3,130 (75.9% fewer). The weather demo on the site goes from 1,727 to 478. Method and files are in the repo; pnpm bench reproduces them. Try it in two minutes: npm create terse-app@latest my-app. The starter ships an AGENTS.md / CLAUDE.md so Claude Code, Cursor or Codex can pick up the language straight away. It's v0 and the vocabulary is small on purpose. What I'd most like to hear is where it breaks: a screen you couldn't express, an error message that didn't help your agent, or the adapter you want next (Material UI is the obvious first one).

Terse

Build React UIs with 75% fewer agent tokens

Build React UIs with 75% fewer agent tokens