Exploring prompt engineering for grapevine pruning decision making with limited data

Abstract

To address the challenges posed by the shortage of agricultural labour, this paper explores the use of Prompt Engineering (PE) to guide Large Language Models (LLMs) for grapevine pruning decision-making. Based on real grapevine images and structured data, five prompt templates were designed and tested on four mainstream LLMs: Gemma-3, Llama-4, Gemini 2.5 Pro, and GPT-o3. Experimental results show that some models excel in consistency—GPT-o3 achieved full consistency (1.0) under Prompt 3, and Gemma-3 performed well under Prompts 2 and 3 (0.98 and 0.96). However, inter-model agreement remained low (Fleiss’ Kappa < 0.07) across all conditions. These findings confirm the feasibility of using PE to simulate expert pruning in data-scarce environments and offer insights for building scalable, interpretable decision-support systems in agriculture.