GPT Prompt revision for kids' magic pen pal

About Magie Post
Magie Post is a Dutch start-up that writes interactive letters to children. Children and their caregivers write to Bibi the flower fairy, who lives in the magical world of Twinkelrijk together with her bunny and other friends. When children share something about themselves, Bibi responds in a positive and encouraging way. In the second part of each letter, Bibi tells about her own adventures in Twinkelrijk and asks the child for help by crafting magical objects that support her in the story.

The adventure follows a standard structure of three to nine letters. However, the beginning of each letter is custom-made and uniquely responds to the child’s own input, for example what happened at school or during the weekend. One of OpenAI’s GPT models is prompted to act as Bibi and write that personalized response.

Project
The previous setup had several problems. For example, Bibi did not always answer all questions from the child, and GPT sometimes closed the letter even though the prompt explicitly instructed it not to, because the generated text was followed by a predetermined adventure continuation.

We solved these issues by moving from GPT-3.5 to GPT-4o. This model was better able to respond to all the child’s questions. The issue of premature letter closing was addressed by forcing GPT to produce an introduction, a response to the child, and a closure in separate JSON fields. This allowed the predetermined adventure text to be inserted between the response and the closure.

Another issue was that children sometimes mention very specific things GPT may not know, such as the name of a restaurant or person. For example, with the term “Monkey Town”, GPT-4o often correctly inferred it was a children’s playground, but sometimes interpreted it as a zoo with monkeys. To avoid such hallucinations, we added a prompt instruction telling Bibi to respond with curiosity and ask what it is when she is uncertain. This also has the advantage of encouraging more natural conversation between the child and Bibi.

Outlook
After revising the prompt to meet the product requirements, we provided Magie Post with several directions for further development:

• Integrating client feedback into the GPT responses after launch.
• Providing GPT with the full conversation history so it can take previous exchanges into account over multiple letters.
• Exploring whether pictures of children’s crafts could be included and interpreted by GPT to create more tailored responses.
• Using OpenAI’s Batch API to lower costs, since instant responses are not necessary in this product context.