Presentation slides are a convenient medium for showcasing ideas, analysis, and pitches. A good slide should capture the creator’s intent, while being aesthetically pleasing and stylistically rich. Recent efforts in generative modeling focus on the former, while the latter is largely underexplored. We propose SlideStyler, which generates slides that are not only consistent with user inputs, but also are visually appealing and presentable. We pose stylized slide generation as Coactive Learning over synthetic preference pairs from SyntheticSlides, our agentic framework emulating iterative human design. Since the classical α-informative oracle needed in Coactive Learning is unattainable for aesthetic tasks, we replace it with a generator-validator composition and prove reward validation preserves α-informativeness under interpretable conditions on each component, thus providing a theoretical basis for the learning algorithm. We perform exhaustive qualitative and quantitative analysis on SlidesBench benchmark, where we see a substantial improvement of 21.25% over the current state-of-the-art method, showcasing the efficacy of Coactive Learning algorithm to make design changes that a typical human creator would perform, such as fixing alignments, repositioning of images, encapsulating bullet points within a background etc., all inferred through learning.