Large Language Models (LLMs) are increasingly used to augment data samples, including time series, despite being designed to process discrete textual data. They operate on symbolic tokens such as words or characters mapped to vectors in a high-dimensional space. With their continuous nature and complex temporal dependencies, time series are challenging to adapt to symbolic sequences. This paper investigates the potential of LLMs to generate synthetic samples of time-series sensor data using only a few-shot learning (3 samples) without fine-tuning. We aim to evaluate their viability in augmenting datasets with a small number of samples, addressing data scarcity and class imbalance challenges. We use Human Activity Recognition (HAR) tasks from wearable device sensors as a case study. Our findings demonstrate that LLMs can produce high-quality synthetic samples in less imbalanced datasets, achieving competitive results compared to traditional generative models for the same tasks. However, the LLM performance decreases with more imbalanced datasets, where the generated synthetic data lacks diversity. We also observed that models trained with LLM-generated samples showed more stability regarding confidence intervals. We also present a framework for evaluating synthetic data generation methods, which can show the trade-off between synthetic and real-world data and suggest practical directions addressing data scarcity and balance limitations.