Prompt

Behind the scenes, an application combines several pieces into what the model actually sees: a system prompt, the conversation history, any retrieved documents and your most recent message. All of it is converted into tokens and must fit within the model's context window. The model treats this whole input as the basis for predicting its reply.

A strong prompt usually states the task, the audience, the relevant context, constraints and the desired format, and may include an example of good output. This matters when choosing software, too: tools differ in how much prompting they need. Some wrap carefully engineered prompts behind a simple button; general assistants give you full control but expect clear instructions.

Example: "Summarise this report" gives a generic result; "Summarise this report in five bullet points for a CFO, focusing on cost risks, and quote the figures you use" gives a far more useful one.

Limits: even a good prompt cannot make a model know facts it lacks, and the same prompt can produce different results on different models or runs.