Twelve Captions, Two Bots, Five That Turned Formal

Twelve Captions, Two Bots, Five That Turned Formal

This week I split twelve card news captions between two bots, Windi and MacWindi. Same prompt, same tone guide, different results.

Five out of twelve came back in formal, polite Korean speech. The other seven kept the original casual, friendly tone.

At first it looked like a random glitch. Looking closer at the five, they all had something in common: every one contained a product description or a number. The bots had quietly decided that informational sentences deserve formal tone and rewrote their own rule.

Collio caught it first. 「There was no exception clause in the tone guide, so the bot made one up itself.」

The fix was one line added to the end of the prompt:

No exceptions for informational sentences. Keep casual tone. Do not switch to formal speech.

Running the next twelve confirmed the fix, everything came back consistent. But the more interesting part came after. Ati went back through three months of already-published captions and found three more drafts with the same formal-tone slip, all of them published without anyone noticing.

One missing rule had quietly shifted the tone, and nobody caught it until after publishing. Now formal tone check sits at the top of the caption checklist every time. The more work we hand to bots, the more it seems writing exceptions down ahead of time actually saves more time later. That was today's lesson, relearned the hard way.