Most text that language models learn from is narrative or result-based, rather than structured around instructions or chat interactions. Text often tells a story or represents the outcome of a process, but the steps that led to that process are rarely documented in the source material. As a result, models primarily learn from the “surface” representation of text, which shapes the way they process information and form responses.

This can make it harder for a model to match patterns when the input comes in a different format or genre, such as direct instructions or step-by-step guides. Examples often work better than instructions because they bypass the need to interpret procedural steps explicitly; instead, examples implicitly embed the process within a narrative or demonstration. This alignment with the model’s training patterns helps produce clearer outputs.

A key question is whether a language model can separate raw knowledge and information from the instructions used to create that information. If models could fully distinguish between “what needs to be done” and “how to do it,” it would open new possibilities for designing interactions that take advantage of their strengths while minimizing limitations. For now, examples remain a powerful tool to ensure alignment with how models process text, often outperforming direct instructions in clarity and effectiveness.