About Killer Island and REAP: We Reap What Crawls

Both games build on a foundation of teamwork to progress through the game and a tense atmosphere, where the slightest textual inaccuracy may result in the loss of the desired effect.

Killer Island

Killer Island is a co-op deduction survival horror game for 3-8 players, set on a fog-covered island where one among you hides a deadly secret. Players must chop wood to stay warm, hunt animals for food, and collect mysterious artifacts for sacrifice — all while trying to uncover the hidden killer before it’s too late. With the voice chat, every whisper, lie, and accusation can change the outcome. Each short, replayable match becomes a tense battle of trust and betrayal, where cooperation is vital — but no one is truly safe.

Killer Island. Source

REAP: We Reap What Crawls

REAP is a co-op horror dungeon crawler with creature collection and farming elements. The player controls a hunter who fights hordes of monsters in dungeons. Procedurally generated levels, the variety of weapon types, and characters’ unique abilities create a high replay value, or level of replayability. Every new run depends on the player’s attention to detail, and their reaction time is put to a serious test, while the atmospheric visual effects and tense soundtrack intensify the immersive effect.

REAP: We Reap What Crawls. Source

The team at Tbjbu2 approached us in the summer of 2025 with the task of translating Killer Island into three languages: Chinese (Simplified), Portuguese (Brazilian), and Russian. Shortly after, we received a request to work on a second project: translating the REAP texts into Spanish.

The developers considered the option of machine translation. However, the texts in co-op horror are extremely sensitive and context-specific: the right terms, consistency, and preserving the atmosphere are important here. Mistakes in the interface, hints, or tutorials can also be irritating in a single-player game. But in a multiplayer, they become critical: players stop understanding each other, and bugs in the UI literally obstruct gameplay.

We discussed the risks of machine translation with the team of developers, and they took the decision to complete the translation with professional linguists in order to preserve quality.

How We Organized Workflow

We received source texts and discussed the estimate and deadlines with the Tbjbu2 team. After that, work on the translation followed the standard pipeline:

1. We selected the translators: native speakers of target languages with a proven track record working on game projects in our database of approved specialists.
2. The resulting team studied the game’s text, and any questions that arose among the translators were recorded in a shared document.
3. We forwarded answers provided by developers to the translators and completed the translation on time, meeting the set deadlines.
4. The final stage of the workflow involved editing and proofreading. We also made use of AI to find semantic errors in the translation.

We paid particular attention to the technical side. We received source files in the .po (Portable Object) file format, and the game developers asked us to exclude certain lines or strings from the translation according to ID. That’s no simple task in this file format: the strings weren’t sorted into columns, and marking them in the CAT program turned out to be impossible.

Our specialists converted the files to Excel, sorted the strings, and then returned the resulting translation to the original file format. Thanks to this system, the developers received a file that was ready to use, fully adapted to their processes.

Going the Extra Mile

Apart from the strings we needed to translate, the file provided by the developers also contained other texts. We decided to run an analysis and check the localization quality of all the material we received in order to help the team at Tbjbu2 avoid errors.

We encountered problems in the text that are typically introduced by machine translation:

• literal word-for-word translations that distort the sense
• different equivalents of the same term
• misplaced variables
• incoherence between parts of a sentence in long segments.

We put together a detailed report and shared it with the team. This type of audit helps developers assess risks and plan work going forward.

Mistakes We Helped to Prevent

To show why machine translation would have been a risky choice in this project, we tested texts in MT. Below, you can see some telling examples of how text might look if AI had worked on it instead of professional linguists.

1. Buttons and controls

AlthoughAI is capable of factoring in context when translating, it introduces some semantic mistakes if the source word can have a number of meanings. For example, AI botched the translation of the “Host” button in this project (in the context of a cooperative game), and translated it as a noun instead of an infinitive verb, even though the key “button_play” could be seen in the string’s ID:

2. Game terms

A funny mistake was made when translating the word “pool”, which AI interpreted literally. This is despite the fact that the literal meaning would be unlikely in word combinations like “energy pool” and “mana pool”:

3. Gender neutrality

A frequent problem is AI’s inability to ensure gender neutrality. If the source sentence is passive and doesn’t have a subject (or if the subject is generalized or variable), and the predicate uses the past participle “-ed” ending (e.g., used, obtained), MT engines translate these words as active and masculine. A human would rephrase to make it sound as neutral as possible:

4. Addressing the player

Another problem encountered in strings was how the player is addressed. AI selects the most direct equivalent, which is usually in the masculine form in translation:

Such errors may seem minor, but when there are a lot of them, they ruin the atmosphere and make a game less smooth for players.