How We're Translating Spirefall into Simplified Chinese

Localization is usually the part of game development nobody blogs about — a spreadsheet disappears into an agency and comes back six weeks later. We're doing it differently, and if you've ever compared card text across languages in your favorite TCG, we think you'll enjoy the details.

The big idea: a machine drafts, a native card player has the final word

Every piece of English text in Spirefall — card names, rules text, those punchy all-caps ability flashes, keyword reminders — lives in a master catalog, like a card database for words. A translation model running on our own hardware (no cloud service, nothing leaves the building) takes the first pass at each line.

But here's the rule everything else hangs on: machine output can never ship. A machine translation lands as a draft, and only text that a native Chinese-speaking reviewer — a card player, not just a translator — has personally approved makes it into the game. Think of the model as an intern writing proxy text, and the reviewer as the head judge who signs off on the official printing.

The glossary comes first, like a set's style guide

Anyone who has read Magic cards in two languages knows the pain of a keyword translated three different ways across one set. So before bulk translation starts, we ratify a glossary with our reviewer: Sin is always 罪孽, Harrowing is always 骇人, Bulwark is always 壁垒, Unit is always 单位. The model is forced to use these exact terms, and any draft that violates the glossary is automatically rejected before a human ever sees it.

This process already caught a wonderfully card-game-specific issue. Chinese card games don't really use the physical "counter" concept the way Western TCGs do. Our reviewer told us entity counters should read as stacks (层) instead — so the usual Chinese renderings of "counter" (指示物, 计数器) are outright forbidden terms in our pipeline. Dozens of existing drafts that used them were purged and redone. Meanwhile "counter" the verb, as in countering a spell, is a completely different word (反制). That's the kind of nuance no machine catches on its own.

Review happens card by card, in context

Our reviewer works from a spreadsheet — but every row carries a link to a private review site that shows the actual card: art, full rules text, English and 中文 side by side, with the exact line under review highlighted. There's even a "try it out" button that starts a match against a practice bot with that card guaranteed in the opening hand, because you can't judge a card's translation without knowing how it plays.

For every line, the reviewer gives a verdict — good, or fix with their suggested wording. Approved lines become shippable text. Suggested fixes become human corrections stored in a protected layer that survives even if we regenerate every machine draft from scratch. Human work is never overwritten. Ever.

There's also a visual proofing gallery that renders the approved Chinese text on the real game components — the card in your hand, the deployed battlefield tile, the ability flash overlay — because Chinese characters at card-frame size need proofing just like a physical print sheet does.

Translation made the English cards better too

A fun side effect: translation is the closest reading a card ever gets. The Chinese review rounds have caught real bugs in the English cards:

• Monument to Tharun said "target Fortress" — but the effect can only ever hit the enemy fortress. It now says so.
• Chains of the Sinful's ability flash read "ATK+1 TO ALL SINFUL," which sounds like a buff. It's damage. It now reads "STRIKE ALL SINFUL."
• Poison their Food turned out to silently do nothing at all — it looked for a card type that doesn't exist in the game. Fixed.
• Bare keyword grants like "Renquist has Harrowing" now carry inline reminder text, so a keyword can't be mistaken for a unique ability — a lesson trading card games learned decades ago, relearned through a reviewer who flagged a "naming collision" that turned out to be the same keyword translated two ways.

The shipping bar

For Simplified Chinese to appear in our Steam Next Fest demo this October, every card and rules line in the demo set must be native-approved — no machine drafts, no stale text where the English changed after translation. The reasoning is very card-game-savvy: Steam shows review scores per language, and a sloppy Chinese card pool would earn worse reviews than shipping English-only.

If the bar isn't met by our freeze date, the demo ships English-only and every approved line carries forward to the full release — no work is thrown away. So far a full faction has gone through two complete review rounds and the glossary is formally ratified. The pipeline works.

More languages are on the roadmap after the fest — Korean and Japanese are next in line, and they inherit everything we've built here: glossary first, native reviewer first, one hundred percent human approval on card text.