Prompt engineering guide: getting better results from AI

A prompt is the instruction and context you give an AI model. The same model can produce very different results from a well-written and a poorly written prompt. Prompt engineering is designing prompts systematically to get consistent, high-quality output.

Structure of a good prompt

1. Role and goal: "You are an assistant helping small businesses with accounting."
2. Task: state clearly and unambiguously what you want.
3. Context: the information, documents or data needed
4. Constraints: length, tone, language, topics to avoid
5. Output format: a structured format such as bullets, a table or JSON

Effective techniques

• Examples: a few input-output examples help the model grasp expectations.
• Step-by-step reasoning: for complex tasks, ask the model to analyse first, then answer.
• Splitting tasks: break a big job into linked small steps.
• Allowing uncertainty: "If the information is not in the context, say so" reduces made-up answers; see our hallucination guide.
• Positive instructions: "do this" usually works better than "don't do that".

Managing prompts in products

• Keep prompts separate from code, under version control.
• Test every change against an evaluation set; see our model evaluation guide.
• Clearly separate user input from system instructions to reduce prompt injection risk; see our AI security guide.
• Avoid needlessly long prompts; token cost and speed suffer.

When it is not enough

If prompting falls short, consider RAG for current or company-specific knowledge, or fine-tuning for consistent style or task performance.

Conclusion

Writing good prompts is the same skill as clear thinking and good briefing. Building it in your product team directly improves the quality of AI features.