Context & personas

Dialecto grounds every machine-translation suggestion and every quality-check evaluation in curated context: the string’s domain, where it’s used, your brand voice, and the audience you’re addressing — and never more of your content than you chose to give it.

The brand-voice profile

Each repo can hold a brand-voice profile: a compact, structured description of how your product talks.

Identity — what the product is, in a line, and its industry.
Audience — who reads these strings.
Tone — four dials: formality, warmth, playfulness, technicality.
Address forms per locale — the formal/informal decision (du vs Sie, tu vs usted), made explicitly for each target locale.
Lexicon — preferred and avoided terms, with notes.
Exemplars — up to a handful of canonical on-voice strings.

Show Dialecto your voice

You don’t fill that in from a blank form. On the repo’s Voice page, show Dialecto material you already have:

paste on-voice copy,
upload a file — plain text, Markdown, HTML, PDF, or DOCX,
or point it at your website and let it read a few pages (it respects robots.txt).

Dialecto distills a draft profile from the material — on the local model by default, so your content isn’t shipped to a third party — and you review each proposed field, correcting and accepting. The manual controls stay available for refining, or for building a profile from scratch.

Audience personas

One product often speaks to several audiences — a formal enterprise admin and a casual end user, say. A repo can hold multiple named profiles (personas), each with its own tone, address forms, and lexicon, with one marked as the default. In-app suggestion and quality-check flows ground on the default profile; the MCP get_voice_card tool takes a persona argument, so editor and agent flows can request any persona’s card by name. Routing personas deeper into the in-app flows is where this is headed.

The voice card

A profile never goes to a model raw. It compiles into a voice card — a deterministic, plain-text block with a hard token budget: identity, audience, the tone line, the address form for this locale, lexicon, exemplars. When space runs short, the lowest-value content is trimmed first. The rendered card is previewable in the app, so you can read exactly what the model will be told.

The card grounds three consumers:

Machine-translation suggestions — alongside the entry’s domain, usage snippet, glossary terms, and translation-memory matches.
The quality check — translations are judged against your stated voice (an informal du where your profile says Sie is a finding).
The MCP get_voice_card tool — your editor or agent receives the same surgical context the in-app aids do. See API & MCP.

Skippable, incremental

Zero context still works: suggestions run generic, and the deterministic checks still run. Every field you add sharpens the output; nothing about the profile is required.