GATE is a learning framework that uses language models to guide task specification and the inference of expected behaviors through free-form, language-based interactions with users. It has been researched in three areas: email verification, content recommendation, and moral reasoning. In a pre-registered experiment, we found that prompting language models executed by GATE with open-ended questions or synthetically generated rich boundary cases often yielded more informative results than user-written prompts or labels. Users reported that interactive task guidance required less effort compared to prompt or example labeling, and provided novel considerations not initially anticipated by the user. Our findings suggest that language model-based guidance can be a powerful tool for aligning models with complex human preferences and values.