Designing a useful prompt

Write prompts that specify a task, relevant context, and the expected response format. Compare instructions with examples, then test whether the model follows them across several inputs.

The earlier lessons explained inference and training. We now turn to application design, starting with the instructions and context sent to a model.

Prompt engineering means designing and testing those inputs for a particular task. A useful prompt reduces ambiguity about what the response should contain. It influences model behavior, but it does not enforce rules with the certainty of application code.

Instructions and decoding settings

The prompt supplies the task and context. Generation parameters such as temperature and top_p separately affect how tokens are selected from the model’s distribution. Changing the prompt and changing sampling are different interventions, so test them separately when diagnosing a failure.

System prompts

Many model interfaces provide a system-instruction field for application-level guidance. The exact roles and priority rules depend on the interface. Use that field for recurring instructions, while the user message supplies the immediate request.