Localization prompts: reconcile terminology and tone

Last updated: 5 Oct 202611 min read

tutorial

A glossary is a decision record, not a command to replace every occurrence of a word mechanically. Give each term its concept, product context, approved rendering, prohibited alternatives, and version. Ask the model to flag inconsistencies and awkward register, then route disputed cases to a native language reviewer with the screen and action attached. A support notice, error, and marketing line may require different tone even when they share terminology. The reviewer should approve the meaning and user action, not merely grammatical smoothness.

Operational case

Harbor uses one approved term for a delivery time range in confirmation notices. A candidate for a failed-delivery message uses a word that instead suggests storage duration. The model flags the glossary conflict; the language reviewer checks the screen and adjusts the message so the recipient understands the next action. The product owner updates the glossary decision for future keys. HP-07 remains blocked until its placeholder and meaning both pass.

Concept: selectable delivery time range
Catalog: H-47 | affected: HP-07 and related notices
Model output: terminology conflict flagged
Human decision: approved wording with screen context
Glossary version: recorded before release

Performance and review cost

Comparing N strings to a glossary of indexed terms is O(N) passes plus the cost of disputed-case review. False positives occur when one surface uses a term in a different sense, so a human decides those cases. The glossary reduces future rework only if each approved change is versioned and propagated to all affected strings.

Common Mistakes

Do not enforce a glossary term without its concept and context.
Do not treat fluent grammar as proof of correct product meaning.
Do not close a dispute without recording the reviewer decision.

Connected lessons

prompt engineering

localization quality