This study proposes a text-guided multiscale topology optimization (MSTO) framework that simultaneously optimizes macro topology and microstructural anisotropy from design intent. Conventional topology optimization (TO) operates on a single density field, which provides no additional design degree of freedom to compensate for performance loss when aesthetic or geometric constraints are imposed. This limitation results in an inherent trade-off between design intent and structural performance. To address this problem, we introduce anisotropic triply periodic minimal surface (TPMS) microstructures, specifically the Gyroid family, as a secondary design variable that independently controls directional stiffness through axis-wise period ratios. The proposed framework consists of three phases. In the offline phase, finite element analysis (FEA)-based MSTO is performed across multiple load cases to generate training pairs of optimized density and anisotropy fields. Each result is projected onto orthogonal planes and annotated with structured captions that separately describe geometry and physics attributes. These image-caption pairs are used to fine-tune a latent diffusion model via low-rank adaptation (LoRA). In the design exploration phase, a designer provides a geometry prompt describing the desired macro shape and a physics prompt specifying load conditions and preferred stiffness directions. The fine-tuned model generates multi-view candidate fields, from which the designer selects a preferred set; a field reconstruction step then recovers a consistent three-dimensional initialization. In the optimization phase, FEA-based MSTO refines the structure from this physically meaningful initialization, updating density via optimality criteria and anisotropy parameters via gradient descent with sensitivities obtained through numerical homogenization. Experiments on a cantilever beam domain confirm that the generative model produces distinct initial fields in response to varying geometry and physics prompts. The subsequent MSTO converges to co-optimized structures where macro topology reflects design intent and microstructural anisotropy recovers structural performance. These results establish that coupling text-guided generative initialization with anisotropy-aware MSTO provides a systematic approach to intent-driven multiscale structural design.