Draft and adapt business material
How to Prepare Game Strings for AI Translation and Approve Them In-Game
Turn game dialogue and UI strings into reviewable AI translation drafts, then check meaning, placeholders and fit in the game.

AI can produce a useful first draft of game text, but the draft should not be the file that ships. To make AI localization for games reviewable, give each string its gameplay context and technical constraints, keep the draft tied to its source ID, have a target-language reviewer resolve meaning and voice, and test the approved export in the game. The handoff is complete only when the text works in its actual scene.
Start with the gate, not a list of words
Imagine a gate scene with two source strings. A stern gatekeeper says, “Wait, {player_name}. The gate closes at dusk,” and a nearby button says “Open gate.”
In isolation, “Open gate” could be read as a command, a description of an already open gate, or an action available to the player. The dialogue needs its own context: the gatekeeper is addressing the player, and “closes at dusk” describes when the gate closes, not a permanent closure. A translator also needs to know whether the button has room for the natural target-language action. The character line and the button share a scene, but they do different jobs.
Make a record that can travel with each string
Before requesting drafts, create one record per string. Keep the fields even if the team stores them in a spreadsheet rather than a localization system:
| Field | Gatekeeper dialogue | Button |
|---|---|---|
| Source ID | gatekeeper.wait_dusk |
gate.open_action |
| Source text | Wait, {player_name}. The gate closes at dusk. |
Open gate |
| Target language | The language and locale being reviewed | The same language and locale |
| Speaker or UI location | Stern gatekeeper, gate scene | Action button beside the gate |
| Gameplay meaning | The gate will close at dusk | Select to open the gate |
| Scene reference | Screenshot or playable scene showing the exchange | Screenshot showing the button in place |
| Approved terminology | The game's approved term for the gate, if one exists | The same relevant term |
| Protected content | {player_name}; record any source tags |
Record any source tags or placeholders |
| Displayed-text limit | Record one if the dialogue UI has it | Record the button's limit if one exists |
These are hypothetical IDs and records, not a prescribed file format. If a screenshot cannot show when the gate closes or what pressing the button does, write that meaning in the record. Do not make a reviewer infer a game rule from a still image.
Crowdin describes screenshots as a way to show where text appears, while Altagram describes using glossaries and context such as tone and purpose. Those inputs can inform a draft. Neither description proves that a particular translation will convey the right rule or fit its UI space.
Draft the line and the button separately
Send the two records as separate translation tasks so each output can be checked against its own purpose. A request for the dialogue could read:
Draft a translation of source ID gatekeeper.wait_dusk for the specified target language and locale. The speaker is a stern gatekeeper addressing the player. Preserve {player_name} exactly and retain any markup from the source. Use the approved game terms in the attached record. Convey that the gate closes at dusk, not that it closes permanently. If the scene context leaves a translation choice unresolved, flag the question rather than inventing a game rule. Return a draft and any question under this source ID.
For gate.open_action, ask for a short action label that means the player can open the gate. Supply its displayed-text limit if the team has one, and ask the system to flag a natural translation that may not fit rather than silently changing the action. Keep the AI drafts and any questions separate from approved translations.
Some products claim to help protect these details. For example, game-localizer.ai says it automatically detects and preserves variables and tags. Treat that as assistance in drafting, not evidence that {player_name} or every tag survived your exported file. Compare the source and export directly.
Give the reviewer the scene and a way to ask questions
Route each draft with its source ID, record, screenshot or playable scene reference, and any AI uncertainty to a target-language reviewer. The reviewer should check whether the gatekeeper sounds stern without becoming a different character, whether dusk still has the intended gameplay meaning, and whether the button reads as an action. They should use the approved term list, compare placeholders and required tags exactly with the source, and mark a decision: accept, edit, or return for clarification.
A question about the game rule belongs with the game team. If “dusk” corresponds to a specific event or the button is unavailable during part of the scene, the reviewer should get that context before approving a translation. Native-speaker review matters especially for story content, humor, and cultural references, as game-localizer.ai recommends. A reviewer still needs the game context; fluency alone cannot establish what an unseen mechanic does.
If the translated button exceeds its recorded limit, the reviewer can propose a shorter label, but should not sacrifice the action's meaning merely to satisfy a count. Record the edited wording and the reason for the decision under the same source ID.
Approve the export on screen
Export the reviewer-approved strings, then open the tested gate scene in the target language. Check the dialogue and button in their actual UI states: no missing text, unintended truncation, broken substitution for {player_name}, or damaged markup. Press the button and confirm that its label still describes the action the player takes. If the game has more than one relevant display size or platform, test the versions the team intends to release.
A screen check is not redundant with linguistic review. Alocai describes a screen check that flags truncation and missing translations, and Crowdin reports a 50-string English-to-French test in which eight strings still needed linguist attention, including context or integration checks. The reported sample is not a prediction for this project, and a flagging feature does not replace watching the scene.
Approve only the tested export. For each shipped string, retain the source ID, context record, reviewer decision, approved text, and export version. If a later build changes the gate scene or button space, that record tells the team what needs another review instead of leaving them to guess why the translation was written that way.
