Large Language Models (LLMs) are rapidly transforming many fields, and now they are poised to revolutionize Computational Social Science. Imagine being able to analyze online discussions, understand public sentiment, and gain insights into social trends with unprecedented speed and depth. But how do we effectively harness these powerful AI tools for social science research? Recent research explores exactly this, comparing different strategies for using LLMs in tasks like sentiment analysis, hate speech detection, and politeness classification. It turns out there's no one-size-fits-all solution. The research reveals three key best practices: larger models with bigger vocabularies perform better, fine-tuning LLMs on specific social science data boosts accuracy, and simple prompts often underperform compared to prompts enhanced with extra knowledge. Interestingly, integrating external databases (a popular technique) isn't always the magic bullet. It can help with smaller models but delivers diminishing returns with larger ones. Fine-tuning on task-specific datasets emerged as a consistent winner for accuracy, especially using newer, efficient techniques like QLoRA, which make fine-tuning more manageable. The takeaway: for optimal performance in Computational Social Science, focus on using powerful LLMs fine-tuned to the specific domain. This means utilizing the knowledge embedded within the model itself through smarter prompting and adapting it further via task-specific training.