What they did
Panin set out to make game localization cheaper and faster for a friend's studio. The first version was simple: a script built one JSON file per language, a translator subagent got each key with its English text and the glossary, a script checked placeholders and another assembled the CSV. He ran it on one release (24 keys, 17 languages, 408 strings) and compared the output line by line with a professional localizers' version of the same file. From that comparison he rebuilt it as a skill with separate roles: a script gate that stops the run when proper names are missing from the dictionary, a translator that gets the string class, length limit, last release's translation and a per-market style card, a deterministic linter, an editor that first reads the target text without the English and then checks it against the source, and a judge that only lists errors while a script computes the score. He does not name the model or the agent harness.
How it went
Meaning held up well: 3 critical errors in 408 strings, such as an English 'of' left inside Chinese text and 'patch notes' rendered as a sew-on badge in German and French. Form did not. The machine copy kept 46 English em dashes across all 17 languages, used straight quotes in 12, left formats like 2x2 unlocalized, picked informal address in French, Turkish and Korean, and left event names in Latin script. His linter flagged 121 issues on the machine file (3 critical, 11 major) and scored the human version 100 in 14 of 17 languages. He reports no results yet for the rebuilt pipeline, only a projection that drafts should pass the linter after three or four releases.
Worth knowing
He puts about 90% of the defects within reach of a plain script, so run deterministic checks before spending tokens on a model judge. The worst errors came from missing data, not the model: empty glossary cells and no record of formal or informal address per market. A high linter score does not mean the text reads as native, and human review stays for anything scoring below 70.