As people use large language models over time, they often develop their own prompting strategies to get better results. These useful experiences, however, tend to remain scattered and unorganized across different chat sessions, making them hard to reuse systematically.
The paper introduces Promptile, a method that recovers and reuses prompting strategies from LLM interaction histories. By mining past conversations, Promptile aims to capture the tacit knowledge users build up through trial and error, turning it into reusable guidance.
The significance lies in treating interaction logs as a resource for improving future prompting. Rather than starting from scratch each time, users could benefit from their own accumulated expertise, and possibly from patterns shared across users. The abstract does not yet describe experimental results, so the practical effectiveness remains to be seen.