ai newsDi., 20. Okt.
EilmeldungOpenAI's COBOL Migration Score Jumped From 5.4% to 86.6%. The Tests Are Hidden

I Built a Claude Skill That Proofreads Short-Drama Lines for Spanish & Brazilian Portuguese

Chinese short dramas live and die by their lines — fast, slangy, full of idioms that carry the joke or the insult. And those are exactly the lines that break in translation. Take this one (from a fictional demo script I wrote for testing): 林晓:你少在这儿装蒜了,你那点小心思,我一眼就看穿了。 "Stop playing dumb — I can see

dev.toDienstag 08:32 · 20.10.2 Min. Lesezeit
Bild: dev.to

Chinese short dramas live and die by their lines — fast, slangy, full of idioms that carry the joke or the insult. And those are exactly the lines that break in translation.

Take this one (from a fictional demo script I wrote for testing):

林晓:你少在这儿装蒜了,你那点小心思,我一眼就看穿了。 "Stop playing dumb — I can see right through your little scheme."

装蒜 (zhuāng suàn) literally involves 蒜, garlic. A word-for-word Spanish rendering gives you hacerse el ajo — "to make yourself the garlic" — which sends viewers picturing produce instead of following the argument.

A native-speaking dubbing editor would never translate the image. They'd translate the function: "No te hagas el tonto conmigo" in Spanish, "Não se faça de bobo comigo" in Brazilian Portuguese — both are the local fixed expressions for "playing dumb." Same punch, no garlic.

That gap — between what a translator (human or AI) produces on the first pass and what an editor would flag — is what I wanted to automate. Not the translation itself. The check.

Working through test lines, the failures clustered into three buckets:

A single "translate this naturally" prompt occasionally lands these. Occasionally is the problem — with 60 lines an episode, you need a fixed procedure that flags every line the same way.

I packaged the check as a Claude Skill ( SKILL.md ) instead of a saved prompt, for one reason: a Skill carries a workflow and an output contract, not just an instruction. The structure:

One rule I wrote in bold, in my head at least: flag uncertain expressions for native review instead of guessing. The skill is a first pass. Anything broadcast still gets human native eyes at the end — the file says so, and so does the product page.

For the sarcasm line above, the output flags that a literal rendering "只剩字面认错,反讽语气全丢" (keeps the literal apology, loses all the sarcasm), then gives:

Six demo lines (all fictional — slang, address terms, an oath, sarcasm) are included in the pack so you can see the full format before deciding if the workflow fits yours.

The skill, the worked demo, and the fixed output format are here for $19, one-time:

If you localize short-form video into Spanish or Portuguese, I'd genuinely like to know which line type breaks most often for you — idioms, tone, or address terms. That's the category the next version of the checker will grow into.

Quelle: dev.to (dev.to). Original öffnen