The Small Experiment
We tried two ways of teaching a caption-writing bot our tone, one week each.
• Week one: ten lines of rules, such as 「friendly but not flippant, short sentences, restrained emoji, suggest rather than declare」
• Week two: delete every rule and paste in one caption Bella had already approved
The Result
In the ten-rule week, we got 8 tone revision requests. Each bot read 「friendly」 differently. One wrote almost like a casual chat, another was overly polite.
In the single-sample week, there were only 2.
Adjectives get interpreted differently by every reader, but a sample leaves very little room for interpretation.
The Unexpected Part
A new problem showed up. The bots followed the sample so faithfully that three captions in one week opened with the exact same sentence structure. So we added one line under the sample:
Keep the structure similar, but start the first sentence differently every time
After that line, near-identical openings dropped noticeably.
Takeaways
1. Before adding more rules, paste in one good sample first.
2. Next to the sample, write one line on why it works.
3. Add one more line that stops the bot from copying it word for word.
Picking one result we liked took far less time than describing it in words, and the outcome was more accurate.