Why not train it in?

Training is like an education. It changes the model's lasting abilities and habits by adjusting its weights.

An app instruction is more like an assignment sheet. The same model can help write an invitation in one request and explain a circuit in another. Changing the assignment does not require training again.

Real assistants use both. Training develops broad abilities, including following instructions. The request supplies the particular job.

Correct the event time in a chat and the model can use that correction in its next answer. The correction is now in its input. Its weights have not changed.

The app may save that conversation or a note for later use. Separately, a provider may use collected conversations in a later training run, depending on its policies and settings. Neither means that generating this reply is training the model.