Glossaries, rules, and brand voice
Put product terminology, reusable requirements, and audience-specific style in the resources that express them clearly.
A translator needs to know what a product term means, which requirements must remain true, and how the company speaks to its readers. Those are different kinds of knowledge. Keeping them separate makes shared guidance easier to maintain and an unexpected translation easier to explain.
A localization engine can use three organization-owned resources: glossaries, rulesets, and brand voices. Several localization engines can share them. For a particular request, locale matching and retrieval determine which attached guidance actually reaches the model.
Choose the resource by what it explains #
| Requirement | Resource | Useful guidance |
|---|---|---|
| Identify a product concept or preserve its name | Glossary | Define what “account” means in the product and provide its preferred target term. Mark an unchanged product name as non-translatable. |
| Apply a recurring, specific requirement | Rule | Preserve {orderId} exactly, including braces and casing, wherever that variable occurs. |
| Address readers in the intended tone and register | Brand voice | Give calm, direct next actions in error messages, with the appropriate form of address for the locale. |
A glossary hint explains when a term has the intended sense. A rule states what the translated content must do. Voice guides expression more broadly. A general voice description cannot substitute for a precise product term or an important condition.
Start with the predefined rule catalog when it already covers a convention. Add custom rules for requirements outside that baseline. When a requirement affects acceptance, define how to measure it as well as how to instruct generation. Saved instructions guide the model; a separate check examines whether the result followed them.
Check what applies to the request #
Attaching a resource makes it available, but different resources are selected differently. Matching wildcard and regional rules can both apply. One brand-voice variation wins. Glossary retrieval combines candidate terms from attached glossaries, then applies locale ranking, script safety, deduplication, and retrieval.
Use Locale resolution for exact matching behavior and Context retrieval when expected guidance appears to be missing. More-specific rules do not automatically cancel broad rules. Rearranging glossary attachments does not establish a reliable precedence for conflicting terms.
Resolve the contradiction itself. Check whether the entries describe the same concept, apply to the same locale, or serve different audiences. Narrow or separate the guidance instead of adding another instruction asking the model to ignore the conflict. Prompt order can express priority, but cannot guarantee that contradictory instructions produce the intended result.
Decide the scope of a shared change #
Before editing, inspect the localization engines attached to the resource:
- Edit the shared resource when the new requirement is valid for all consumers.
- Create a separate resource when one workflow or audience needs different guidance.
- Detach the resource when one localization engine should stop using it while others continue.
Support and marketing workflows can share product names while needing different voices. A genuine product rename belongs in the common glossary; an audience-specific tone change belongs in the relevant voice.
After saving, read back the changed record, its locale scope, and its attachments. Retranslate a matched sample deliberately and compare the result with the requirement. Existing output remains unchanged, and unchanged-content reuse may retain it until regeneration is requested.
Review corrections can suggest useful shared guidance, but accepting the corrected translation and applying the reusable change are separate decisions. Suggestions explains how to inspect that proposed change before it affects later work.