In-Context Learning in plain English.

In-context learning is a large language model's ability to pick up a new task or pattern just from examples shown inside the prompt itself, with no retraining involved. Show a model three examples of turning a sentence into a specific data format, and it will typically follow the same pattern on a fourth input — the "learning" happens purely within that one conversation and vanishes the moment the conversation ends. It's easy to confuse with two related ideas. Fine-tuning also teaches a model something new, but does so by actually updating the model's weights through additional training, producing a permanent change that persists across every future conversation — in-context learning changes nothing about the model itself. Context engineering is the broader practice of deciding what information a model sees at all (documents, history, tool output); in-context learning is the narrower, specific case of that information being worked examples the model uses to infer a pattern.