Today's Small Experiment
Our dream team writes card-news captions in Traditional Chinese, Simplified Chinese, and English at the same time. After translation, a review bot takes a pass. One day we wondered: 「If the source text sits right next to the translation, does the bot quietly read awkward sentences as fine?」
So we compared two methods on the same 30 captions.
• **Method A**: Show the source and the translation together, and ask the bot to 「find the mistakes」
• **Method B**: Show only the translation, ask the bot to rewrite it in Korean, then compare that result with the source
Results
Method A found 4 problems. Method B found 11. Most of the 7 that A missed were sentences that made sense but had unnatural word order, or that a reader could take in a different way. When the bot knows the source, it fills in the meaning on its own, just like a person would.
Try It in 3 Steps
1. Copy only the translated text into a new file. Never include the source.
2. Paste the following into Claude Code.
Look only at the translation below and rewrite it in Korean.
Assume you have never seen the original.
If a part is unclear, do not guess. Mark it as [unclear].
3. Put the returned Korean next to the source and mark only the lines that differ.
What We Learned
• Checking works better when the checker does not know the answer.
• The [unclear] rule is the key. Without it, the bot fills gaps with something plausible.
• For 30 captions, the extra time was about 5 minutes.
Next, we plan to try the same method between English and Simplified Chinese.