Subtext is a context-aware localization pipeline for indie games and doujin studios — JP↔EN first. The whole script stays in-context, so every string comes out consistent.
The problem
A CSV of dialogue lines tells a translator nothing. Who is speaking? What pronouns do they use? Which item name is canon? Generic MT guesses — and guesses become bugs that ship.
Product
A translation pass that reads like a human localizer's process — because it holds the same context a human would ask for.
The full game script, glossary, and per-character voice sheets stay in-context across the entire pass — not string-by-string.
Per-character register is preserved: politeness level, dialect, gendered speech, catchphrases. A knight doesn't sound like a mascot.
Genuine ambiguities are surfaced to the writer as questions with suggested options — instead of being silently decided.
Imports from CSV, Unity, and Ren'Py. Review UI on top. Exports back in the same format — no reformatting your project.
Pipeline
Designed to slot into an indie team's existing workflow, not replace it.
Script strings from CSV, Unity localization, or Ren'Py, plus your glossary and character sheets.
Claude reads the whole script and voice sheets, building the context a human localizer would need.
Every string is translated in-context. Ambiguities become questions for the writer, with options.
Your team reviews in the UI and exports in the original format, ready to ship.
Why Claude
A game script plus its glossary and voice sheets is exactly the problem frontier long-context models are built for — and JP↔EN nuance (register, keigo, implied subjects) is where smaller models fall apart. Claude is the core of the product, not a bolt-on.
100% of our AI spend is on Anthropic.
Status
We're running pilots with indie developers and doujin circles in Tokyo. If you're shipping a JP↔EN title and tired of fixing MT output, we want to talk.