Generative AI Design Patterns

Chapter 2. Controlling Content Style

The patterns in this chapter all have to do with controlling the style of the content (such as text, images, and video) generated by foundational models, which are models that have been trained on large datasets and are capable of generating a wide variety of content. However, the style of that content will, by default, be based on the training process that the model provider has used. Even if you stick to a single model version, model responses are stochastic—which means you may get different responses even if you repeat a question exactly (assuming that there is no caching going on). This means that downstream applications and end users that use responses from a GenAI model will have to deal with quite a wide variety of possible styles.

For example, we asked a number of foundational models the same question:

What’s a good side dish for pierogi? Answer in a single sentence.

We then recorded the answers (see Table 2-1). As you can see, even when we restrict the answer to a single sentence, different models can answer the same question very differently. Imagine how much more diverse the results would be had we not restricted the style of the response in any way!

Table 2-1. Answers to the same question in different styles (answers retrieved from the models in February 2025)

Model	Model provider	Answer
GPT-4	OpenAI	A great side dish for pierogi is sautéed onions with butter and a sprinkle of crispy bacon bits.
Claude Sonnet 3.5	Anthropic ...