Amandine Caut: Documenting and Improving Prompts for Large Language Models

Abstract

Generative Artificial Intelligence has fundamentally reshaped the landscape of technology, transitioning from academic research into a revolutionary force across virtually every sector of society. Today, generative AI drives innovation across numerous domains, reflecting its expanding societal impact.

At the forefront of this revolution are large language models (LLMs): sophisticated systems that have redefined human-AI interaction and democratized access to intelligent analysis. As LLMs become increasingly accessible, the critical challenge is no longer whether to use them, but how to deploy them effectively and responsibly within specialized domains.

This thesis investigates how Large Language Models can deepen our understanding of the world and enhance our capacity for insight and discovery. Specifically, it examines the practical integration of LLMs into data science workflows, exploring how they can enhance data exploration, strengthen analytical reasoning, and support informed decision-making. Equally important, it addresses a fundamental challenge often overlooked: establishing rigorous documentation and structural practices that ensure reproducibility, transparency, and methodological integrity when working with these powerful tools.

By synthesizing theoretical foundations with practical application, this research provides both technical insights and actionable guidance for LLMs into data science practice: not as a shortcut, but as a disciplined, transparent, and accountable approach to modern analytical work.