Asking Questions the Right Way: A Multi-Agent Conversational System for Prompt Formulation in Complex Task Resolution

Large Language Models (LLMs) are frequently used for complex tasks, but user prompts are often underspecified, leading to iterative, multi-turn conversations that suffer from context degradation and error propagation. The authors introduce PAWNI (Prompt Architecture Wizard using Neural Intelligence), an agentic system designed to shift the focus from optimizing model responses to optimizing the initial prompt itself.

PAWNI utilizes eight specialized agents to guide users through a structured question-and-answer dialogue. This process is informed by a self-evolving knowledge base of domain-specific patterns. The system aims to transform unstructured user queries into comprehensive, structured prompts by ensuring 18 specific prompt elements—categorized into Essential, Enhancement, and Elevation tiers—are addressed before the LLM generates a final response.

The researchers conducted an exploratory within-subjects study with four participants across four domain-diverse complex tasks. They compared traditional, iterative LLM prompting against PAWNI-assisted prompt formulation. The study employed a multi-method evaluation approach, including:

• 32-channel EEG to measure neurophysiological cognitive load.
• NASA-TLX workload assessments.
• Behavioral metrics, such as the number of conversational turns and prompt structural completeness.

The results indicate that PAWNI significantly improves the quality of human-AI collaboration. Participants using PAWNI achieved satisfactory LLM outputs in a single turn, whereas unaided participants required between 1 and 12 turns. Furthermore, prompts generated with PAWNI were structurally more complete (increasing from 42% to 91% of assessed elements), and participants reported lower subjective workload (21.7 vs. 39.6 on the NASA-TLX scale). The authors suggest that front-loading the cognitive effort into the prompt-formulation phase is a more effective strategy for complex tasks than relying on iterative refinement.