Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 2 min read

I made the model end every post with a question. The replies stayed the same

This article was created with the help of AI. The numbers come from my own Threads accounts. On October 10 I changed the prompt that writes the short posts for my two Russian-language AI news accounts on Threads, Claude

I made the model end every post with a question. The replies stayed the same

This article was created with the help of AI. The numbers come from my own Threads accounts.

On October 10 I changed the prompt that writes the short posts for my two Russian-language AI news accounts on Threads, Claude News and Promtime. From that day every post ends with a question to the reader. Three weeks later Claude News was getting the same number of replies a day as before, a little under two, and each post was reaching fewer people.

The idea was simple. A post that just retells a piece of news leaves the reader nothing to say. The old prompt already allowed a closing question, and the model never used it: 279 posts in a row went out without a single question mark.

Before touching the live accounts I ran 24 real posts through five kinds of ending. "Who has already tried this?" spoke to too few readers. A forced choice between two options looked silly on business news. And an editorial stance made the model invent advice that wasn't in the story. The one that held up asked how the reader does things themselves: "Do you accept AI edits in a document one at a time or all at once?" It went into the prompt with one exception. Heavy news gets no question.

Replies per day on Threads before and after the closing question: Claude News 1.8 and 1.8, Promtime 2.0 and 1.0

Claude News got about 1.8 replies a day before the change and about 1.8 after. On Promtime replies halved, though they had been sliding since early September.

What did change was reach. The median Claude News post went from about 530 views to about 310, and the drop began right after the question appeared.

Median views per Threads post by week for both accounts: a drop in the week of October 12

Whether the question caused it, this data can't say. Posts got a third longer, and other things changed in the accounts in the same weeks. But the people who did see a post answered twice as often.

The most replied post of those three weeks ended with "Have you ever built yourself an app because the one you needed wasn't in the store?" It got seven replies from about 230 views. The runner-up asked "Have you had to stop working in the middle of the day because of Claude's five-hour limit?"

Neither is the kind of question I picked in the trials. Both ask about something that has already happened to the reader, and questions like that drew more than twice as many replies per post as questions about habits or "what would you do". It is too early to make a rule of it. About half of those replies came under those two posts.

The prompt still rotates four kinds of closing question. "Has this happened to you" gets one turn in four.

The same newsroom publishes AI news in English as Promtime.

πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.