Prompt engineering and a custom ‘Research Question Builder’ agent, explained

Esmeralda Rogers, an evidence review and synthesis librarian at the Texas A and M University Medical Sciences Library, said a prompt is simply “the question or instruction you put into the LLM,” and that output quality depends on how the prompt is structured. She introduced three frameworks—CLEAR (concise, logical, explicit, adaptive, reflective), CAST (criteria, audience, specifications, testing) and CORE (context, objective, role, example)—and showed side-by-side examples of an unstructured prompt versus a structured CORE prompt that produced more searchable keyword lists.

Rogers demonstrated a Copilot example that initially returned long phrase keywords and then produced a table organized by population, intervention and outcome after using a structured prompt. She showed a custom agent the team built, called Research Question Builder, which pulls only from curated knowledge sources (their LibGuide, workshop slides, and open-access articles) and is intended to generate structured review questions and recommend appropriate frameworks like PICO or PCC.

Why it matters: Rogers argued structured prompting makes LLM outputs more useful for systematic review tasks such as keyword generation and framework mapping. She cautioned about platform costs and privacy: the team moved their agent from a premium-only Copilot environment into Copilot Studio to publish more broadly, but noted potential data-usage and training implications for user inputs.

Practical takeaway: Use structured prompt frameworks for complex queries, pilot any custom agent with restricted knowledge bases, and document restrictions so the agent only performs the permitted tasks. Rogers said resources and references will be included in the webinar materials posted after the session.