I want to use Claude to draft translations for a small app, but these are interface strings rather than ordinary paragraphs. Some contain variables such as {customerName}; others use ICU MessageFormat plural rules.

A translation can sound perfectly natural and still break the screen if it changes a variable name or damages the message syntax. There’s also the less obvious problem of translating a short label without knowing whether it is a button, a status, or a noun.

For anyone using Claude for UI localization, what does your review process look like? I’m particularly interested in placeholder validation and how you give the model enough screen context without sending the entire application.

Would you validate syntax and variable names automatically, then reserve human review for wording and language-specific plural behavior? Or have you found that separating ordinary strings from plural messages produces more reliable drafts?