Prompt Engineering Guide: Techniques That Work in Production

State the task, the audience, the length and the format. Then add the reason: "Keep it under 50 words because it's shown in a phone notification" beats "be short", because the model can make better trade-offs when it knows the goal.

Two to five input→output examples teach format, tone and edge cases better than paragraphs of description. Include at least one tricky example, like an "I don't know" case.

Wrap documents, emails and user input in clear delimiters such as <document>…</document>. It reduces confusion and lets you refer to parts precisely. It is not a security boundary on its own: untrusted text can still contain prompt injections.

If code consumes the output, ask for JSON with named fields, and use your provider's structured-output or schema mode when available. Always validate the result.

For multi-step problems, asking a standard model to think step by step (chain-of-thought, Wei et al. 2022) improves accuracy. Reasoning models do this internally. For them, give clear goals and constraints instead of micro-steps.

Keep a set of real test inputs and run it on every prompt change. Track pass rates. Without evals, "improving" a prompt often quietly breaks other cases.