verba

verba is a writing app you talk to. you dictate prose and issue spoken commands, and an editing model applies the changes directly to your document.

the key idea is that it’s an editing model, not a retrieval one — there’s no search step. it takes the text you have and the command you said, and produces the text you meant.

the design system

everything visual comes from a single set of design tokens (tokens.css) — no hardcoded colors or spacing anywhere in the app. the palette is deliberately quiet: warm paper tones, near-black ink, a sparse blue accent, and red reserved for exactly one thing — signalling that dictation is live. serif type shows up only inside the document you’re editing; everything else is clean sans-serif ui.

there’s one component library, @verba/ui, rendered across three surfaces — a component gallery, the production editor, and storybook. change a token and it propagates everywhere at once, so the mockups and the real app never drift apart.

evaluating it

testing an editing model is awkward because there’s rarely one right answer. verba’s eval set is built from triples: some original text, a spoken command, and the edit i’d expect out the other end. that set doubles as a regression suite and as training data for fine-tuning later.

there’s more on the design system and the eval approach on the blog.