The AI features

Three features run on the AI engine. Each is a row under AI Settings → Routing by feature, each with its own engine preset — one model can serve all three, or any one can point at a stronger model.

Translate

The batch translate loop — the tool’s core. Pending keys go out in batches: each string is shielded first (placeholders, glossary terms and markup are swapped for tokens the model cannot break), sent with your per-language conventions and reviewer notes, then restored and checked on the way back. A string whose shield comes back damaged is refused, never written.

Review

The back-translation pass: the translated string is translated back toward the source and shown beside it, so a reviewer who doesn’t read the target language can spot drift. Shown to the reviewer only — never written to a catalogue. Review deliberately routes to the same engine the translation used; a second opinion from the same model about its own work is the point (what changed in back-translation is signal, not judgment).

Confirm

The second opinion on byte-identical targets — strings whose translation came back identical to the source. One key per call, by design. It annotates (“plausibly identical in this language” vs “looks untranslated”); it never signs off a string by itself.

What is deliberately not here

Extract (docs front-matter → locale keys) has no routing row: it is pure parsing with no engine call anywhere, and a routing row for a feature that cannot route would lie. It re-appears the day it gains a real AI step.

Each feature’s prompt

Each feature has one prompt, and you can read and change it in the AI page’s Lab (Routing by feature → the feature): its instructions, its message, and its JSON answer. The parts in {{double braces}} are filled from your project on every run — the languages, the context sentence and the glossary from Setup, the per-key notes from Review, and the strings themselves, with placeholders and glossary terms already swapped for ⟦0⟧-style tokens. A rule that doesn’t apply to your project (no glossary, no plural separator) is left out whole. Reset to default puts the shipped prompt back.

A run checks its prompt before it starts. A prompt without {{items}} (where the strings go), with Output as JSON off, or using a {{name}} the run doesn’t fill is refused, with the reason and the fix; the confirmation pass checks its own prompt when it starts. See Set up the AI engine for trying a prompt in the Lab.