Two pieces of writing do most of the work in a localization project: a short brief describing the product, and a list of words that must never drift. Both are maintained per project and both reach every translator, human or machine.
A few lines describing the product, its audience and the constraints a translator would otherwise have to guess. Set per project, so a student banking brand and a private bank can share one key set and still sound nothing alike.
Free text, set per project. Different clients can have completely different tone instructions on the same key set.
A custom project starts with the default project's brief. Override it only when a client's tone genuinely differs.
Because the brief travels with each string, length costs you on every key. Four useful lines beat four paragraphs of positioning.
Product names, regulated vocabulary, the word your CEO cares about. Lock it once per project and every locale obeys — AI and humans alike.
A custom project starts with the default project's glossary and brief. Most clients never need their own — the wording that differs is usually a handful of keys, not the whole vocabulary.
When one does, override just the terms that change. A white-label bank calling the wallet a Konto adds one term; everything else keeps reading through.
The glossary is not only an AI input. A reviewer editing a cell sees the terms that apply to that string, with your preferred translation beside each one, so a human and a model reach the same answer.
A string that breaks a locked term is flagged in the grid whether a person or a model wrote it. Approving it anyway is allowed — with a note recorded on the revision.
The brief is sent with every request, so length costs you on every string. Say what the product is, who reads it, and the two or three constraints a translator would otherwise guess wrong.
Retail banking app for Germany. Address the user formally. Never translate Kestrel, IBAN or SEPA. Keep button labels under 20 characters.
Every clause changes an output. A translator could not have inferred any of it from the strings alone.
We are a fast-growing fintech committed to delighting our customers with a seamless, best-in-class experience across all touchpoints.
Nothing here changes a single translation. It costs tokens on every key and dilutes the instructions that matter.
Register belongs on the language, not here — see formality.
Cheaper than correcting four thousand strings that all got it wrong the same way.