A good prompt describes the task and the output you want. The practices below are starting points; a comparison on recurring examples shows whether they work for your tasks.
Eight practices
- 1Name the task: Say what should be created, compared or extracted.
- 2Bound the context: State the relevant audience, purpose and approved sources. Avoid unnecessary confidential data.
- 3Describe the output: Specify language, length, structure and fields where they matter.
- 4Give examples: A short example can clarify format or tone. Check whether it also introduces unwanted patterns.
- 5Name the criteria: Explain how to recognise a useful answer, for example completeness and cited sources.
- 6Allow uncertainty: Ask the model to name missing information instead of filling gaps.
- 7Break tasks down: In multi-step work, separate steps are easier to check.
- 8Iterate with test cases: Compare several prompt versions on the same representative examples and assess errors, not just style.
What prompts cannot do
A prompt guarantees neither correct facts nor higher productivity; results still need human review. Role instructions add context, but they do not give the model real professional experience or additional facts. Likewise, “think step by step” is not a reliable way to force correct conclusions; do not treat the reasoning a model outputs as proof. Whether a prompt saves time only becomes clear when you measure it in the actual workflow.
Maintaining a prompt library
For a prompt library, record the task, model, date, test cases, known limitations and the person responsible. Remove outdated examples when the model or the process changes.
Sources: OpenAI: Prompt engineering · Microsoft Learn: Prompt engineering for Azure OpenAI

