Micro-gesture recognition is an essential research task for various applications. The inherent subtlety of micro-gestures and their susceptibility to noise make them challenging to detect and classify reliably, especially under complex, unconstrained environments. Existing single-modality approaches struggle to capture comprehensive spatio-temporal patterns or suffer from robustness limitations, while current multi-modal methods often fail to fully leverage complementary information due to ineffective cross-modal feature alignment. To address these challenges, we propose RGB-Text-Skeleton Union (RTS-Union), a novel multi-modal framework that integrates RGB, skeleton, and text modalities to extract fine-grained gesture representations and achieve robust micro-gesture recognition. Specifically, we introduce three specialized modules: (1) a MLLM-based fine-grained text enhancement (MFTE) module that enriches gesture labels with detailed semantic descriptions; (2) a lightweight spatio-temporal RGB enhancement (LSTRE) module that efficiently captures subtle motion cues using a CLIP-based adapter scheme; and (3) an explicit skeleton extraction (ESE) module that generates discriminative skeleton features from pseudo heatmaps. A contrastive learning strategy aligns RGB and skeleton features with the enhanced text representations, encouraging effective cross-modal interaction. Finally, a Kolmogorov-Arnold Network (KAN) classifier models non-linear relationships between modalities for accurate gesture classification. Extensive experiments on two benchmark micro-gesture datasets demonstrate that our approach significantly outperforms existing state-of-the-art methods.