Title : The rhinologist as developer: A practical framework for prompt engineering and app creation in clinical practice
Abstract:
Objectives: Generative AI and Large Language Models (LLMs) have lowered the technical barrier for software creation, yet a gap remains between potential and practical application for surgeons. We aim to bridge this gap by defining a structured "Prompt Engineering Framework" specifically for Rhinologists. Our objective is to demonstrate how clinicians can utilize specific prompting strategies to build NHS-compliant, effectively functioning medical software without prior coding experience.
Methods: We performed a literature review of current prompt engineering methodologies (e.g., Chain-of-Thought, Few-Shot prompting) with experiential data derived from the development of ‘Sniffinsticksonline’—a fully functional clinical web app built using natural language. We deconstructed the development process into an iterative workflow, analysing the specific prompt structures required to overcome common AI pitfalls such as logic errors ("hallucinations") and security compliance issues.
Results: We present a validated, three-stage "Clinician-to-Code" workflow:
Contextual Architecture: How to structure initial "system prompts" to define clinical parameters and safety constraints.
Iterative Refinement: demonstratable success using Chain-of-Thought prompting (asking the LLM to explain its reasoning step-by-step) to debug complex clinical logic, significantly increasing code accuracy (Shah et al., 2024).
Cross-Verification: The utility of using multiple LLMs to audit code for NHS digital compliance. This workflow was successfully used to build ‘Sniffinsticksonline’, which achieved superior usability scores compared to the legacy OLAF system in subsequent validation trials.
Conclusions: Proficiency in prompt engineering is emerging as a critical competency for clinical innovators. We conclude that by applying this structured framework, Rhinologists can rapidly prototype and deploy digital solutions, removing the dependency on external developers. ‘Sniffinsticksonline’ serves as a proof-of-concept, demonstrating that clinician-led, AI-architected software is not only feasible but can outperform existing solutions.
