Customized molecule generation remains a challenging task, especially for data-scarce categories such as atropisomers. Their structural complexity and scarcity of labeled data often lead to unstable optimization and poor controllability in traditional diffusion models. To address this gap, we propose AtropDiff, a diffusion framework guided by a multi-task pretrained classifier that implicitly encodes stereochemical knowledge. Key innovations include: 1) multi-task pretraining of a chirality-classifier for gradient-guided chirality control; 2) dynamically weighted gradient guidance to balance chirality accuracy and functional group validity during denoising; and 3) progressive integration of scaffold into the generation process to enable optimization on known atropisomers. Our results demonstrate that AtropDiff successfully overcomes data limitations, advancing AI-driven scientific discovery in chemistry.