Localization QA for game teams
Dohwa checks translation files before they reach the build: missing placeholders, broken tags, text that overflows the UI, glossary drift, and, with Claude, translations that say the wrong thing.
The checker on this page runs in your browser. Files are not uploaded.
You have {0} gold.
골드를 가지고 있습니다.
Defeat the Ember Wyrm before it reaches the village.
잿불 비룡이 마을에 도달한 후에 처치하세요.
meaning "before" became "after". No rule can catch this; a reviewer that reads both languages can.
Live checker
Load the sample from a fictional RPG (English to Korean, 30 strings) or drop your own CSV or JSON with key, source, target, and optional max_length columns.
Rule checks run locally in this page. Meaning-level review uses the Claude API and is not run from this page.
| Key | Check | Finding | Source / translation |
|---|
Claude review on the sample
The sample also hides six meaning-level errors: two mistranslations, two omissions, and two strings in the wrong register. The Claude review harness for this sample is built; its results will be published here after the first run.
What it catches
Deterministic checks are instant, free, and never miss a broken placeholder. Claude reads source and translation together for what rules cannot judge.
Each finding comes back as structured JSON with the key, the type, an explanation, and a suggested fix, ready for a CI report or a translator's queue.
Who it's for
Localization bugs are cheap to fix in a spreadsheet and expensive to fix after a store review or a patch day.
Shipping five languages with contract translators and no dedicated LQA pass.
New event, store, and patch strings every week, translated on a deadline.
Reviewing many titles at once and needing a first pass before human testers.
Why Claude
A string table is long, multilingual, and full of context that rules cannot hold. Each part of the review maps to a Claude capability.
Strings, glossary, and style guide in one pass, so consistency is judged across the file, not line by line.
Korean speech levels, Japanese politeness, gendered forms: the errors players notice first.
A fixed JSON schema per issue, so results merge with the rule checks and feed CI.
Next: connect to repositories and translation tools so review runs on every change.
Team and roadmap
Dohwa is built by a solo founder covering product and engineering, based in Seoul.
Live on this page. On the sample file it flags all 12 rule-detectable errors with no false alarms on the 12 clean strings.
Sends the table, glossary, and style guide to the Claude API and scores the findings against seeded meaning-level errors.
A command-line tool and GitHub Action that fail the build on broken strings, piloted with indie studios.
No. The checker on this page reads the file in your browser and sends nothing over the network.
CSV and JSON today; XLIFF, PO, and Unity localization tables are next. Rule checks work for any language pair. The Claude review is being evaluated first on English to Korean.
No. It removes the mechanical errors and points reviewers to the strings most likely to be wrong, so human testers spend their time on what needs judgment.