Prompt

A prompt is the input supplied to guide a model's response for a request. In a chat application it can include instructions, the user's question, selected conversation history, documents, and tool results. The visible message box may show only part of what the application supplies.

Make the task and evidence explicit

Suppose the task is answering a return question. “Be helpful” leaves many choices open. A more specific instruction can ask for an answer based on the supplied policy, preserving the opened-item exception and identifying missing information. The policy text is evidence to interpret; the application's trusted instructions define how to handle it.

Clear inputs reduce ambiguity, but they do not prove that the output follows the policy. Compare the answer with the relevant source and task criteria. Hugging Face's generation guide illustrates supplying input text to guide continuation.

Changing input is not changing training

Editing a prompt normally changes the current computation without updating learned model parameters. It also does not automatically update a colleague's separate conversation or the application's shared document store.

The application decides what to retain and include in later requests. It must separately enforce permissions and tool limits. Writing “only access my orders” in a prompt is not a replacement for checking ownership when the order service receives a request.

Reference: Hugging Face: Text generation.