Few-shot action recognition classifies human actions in videos from very few labelled samples. Temporal redundancy and background ambiguity misalign visual features with action semantics, so plain video-text similarity is insufficient. We propose ReAlign, which injects semantics and aligns features adaptively. A cross-modal prototype refinement (CMPR) module uses fine-grained text as a prior to focus support prototypes on action-relevant patterns while suppressing background noise. A text-adaptive query alignment (TAQA) strategy then calibrates query representations against the holistic textual space, reconstructing features by semantic affinity to maximize category discriminability. On five public benchmarks ReAlign outperforms state-of-the-art methods in both 1-shot and 5-shot settings.