Text data augmentation typically involves adding noise to the original text, substituting words based on context, or rephrasing the original sentence. These traditional methods do not offer new perspectives that could help classifiers generalize better to new instances. This, combined with the fact that in certain fields, such as hate speech detection, creating these datasets is a highly intensive task, from collecting positive samples to labeling them, which usually requires experts in the target topic. In this work, we present Low-Resource Dataset Synthetic Generation (LoRDS-GEN), an automatic text data augmentation method based on LLMs (demonstration-based generation), capable of providing varied samples that respect the original writing style. Our method focuses particularly on low-resource hate speech datasets, but we empirically discovered its potential and consistency even in larger and more robust datasets. The design of the prompt, combined with token sampling strategies, positions LoRDS-GEN as the most consistent alternative for LLM-based synthetic data generation compared to other existing methods and prompts in the state-of-the-art. Our benchmark includes (1) CMSB, a sexist dataset, and (2) ETHOS, a highly varied dataset with multiple forms of hate speech. LoRDS-GEN outperforms other traditional augmentation methods such as NLPAug and BackTranslation.