Prompt Optimization for Enhancing AI-Generated Creative Writing with Large Language Models

Abstract

This thesis examines how prompt engineering strategies influence the quality of AI-generated creative writing across multiple large language models. Eight prompt variants were tested on four model families using PromptLab, a purpose-built browser-based research platform. Evaluation followed a dual pipeline combining an LLM-as-judge with deterministic mathematical metrics covering lexical diversity, style match, and coherence. Results show that structured and persona-based prompts outperform zero-shot baselines by up to 36.5%, with Gemini 2.0 Flash leading qualitative scores and Llama-3.2-3B leading lexical diversity metrics. The study confirms that prompt structure has a significant, measurable effect on creative writing quality and that cross-model comparisons are essential, as different model families respond differently to the same strategies.