Researchers unveil PAWNI system for smarter AI prompting

A team led by B. Sankar, Pawni Yadav, and colleagues at an undisclosed institution has developed PAWNI (Prompt Architecture Wizard using Neural Intelligence), an agentic conversational interface designed to optimize how users formulate prompts for large language models (LLMs). Unlike traditional systems that focus on refining model responses, PAWNI prioritizes refining the initial query itself through a guided question-and-answer dialogue, backed by a self-evolving knowledge base. The system deploys eight specialized agents that collaborate to transform unstructured user input into structured prompts using a three-tier framework of 18 elements divided into Essential, Enhancement, and Elevation categories.

In a within-subjects study with N=4 participants tackling complex tasks, PAWNI demonstrated significant improvements over unaided prompting. Participants achieved 42% to 91% more structurally complete prompts, reported 45% lower cognitive workload (39.6 vs. 21.7 NASA-TLX), and consistently produced satisfactory outputs in a single conversational turn—compared to 1-12 turns required without assistance. While the sample size limits statistical robustness, the directionally consistent results suggest that optimizing prompt formulation at the front end could be a critical lever for enhancing human-AI collaboration.