The Day the Blanks Showed Up
We added one line to the dream team's translation bot prompt: 「If you are not sure about an expression, do not invent one. Leave it blank as [NEEDS CHECK].」 At first we worried. If drafts came back full of holes, it seemed the workload would only grow.
What Happened
It went differently than expected. There were only one or two blanks per article, mostly proper nouns, unit formats, and idioms. Before, those spots were filled with plausible guesses, and the reviewer had to read the whole piece with suspicion from the first line. Now they only look where the bot raised a hand.
Why It Worked
• Review now starts from 「check the flagged spots」 instead of 「read everything」.
• When the bot exposes its own doubts, errors have nowhere to hide.
• Traditional Chinese, Simplified Chinese, and English often went blank in the same place, which hinted that the source text itself was ambiguous.
The Recipe
If unsure, mark it as [NEEDS CHECK] and give the reason in one line.
A fixed marker means one search collects every blank. We ran the search on the Windows device to build a list, then had the Mac mini bot propose fixes.
Homework Left
A bot that leaves too many blanks is useless. So we added a second rule: if an article has more than five blanks, fix the source text before translating.
A bot that can say 「I am not sure」 saves more review time than a bot that simply writes well. That is what one small experiment taught us.