An intern asked me one question about my AI chatbot and it fixed the whole thing

Last Thursday in our office in Austin, I was showing a new intern the customer service bot I've been training for weeks. It kept giving these long, robotic answers that nobody seemed to like. She just asked, 'why does it always say as an AI language model?' and I realized I never actually removed that default intro from the system prompt. I spent 3 hours rewriting the instructions to make it sound like a real person, using plain talk and shorter replies. The next round of test chats felt totally different, way more natural. Has anyone else hit this wall where the bot's training data makes it sound too stiff and you have to force it to sound human? I'm curious what tricks you all use to get that conversational tone down.

1 comments

1 Comment

cameronwest16d ago

Ha, that reminds me of my buddy who runs a little support bot for his shop. He spent a month tweaking responses and still got complaints it sounded like a robot. Turned out his whole prompt still had "I'm here to assist" baked in from a template. Once he stripped that and told it to just answer like a coworker, night and day difference. It's always the dumbest default setting that kills you.