Prompt Engineering interview questions

Designing instructions and context so an AI model gives accurate, useful and consistent answers.

A clear task, relevant context, constraints, the output format and examples where helpful.

Zero-shot gives only instructions. Few-shot adds examples of input and output to guide the model.

They separate instructions from data, so the model does not confuse pasted text with your instructions.

<instructions>Summarise the ticket</instructions>
<ticket>...</ticket>

Instructions that set the model's role, rules and style for the whole conversation.

Asking the model to work through the problem before answering, which improves accuracy on multi-step tasks.

Specify the exact JSON shape, give an example, and validate the result in code.

Splitting a big task into smaller prompts where each output feeds the next.

Confident but wrong answers. Give source material, ask it to quote sources, allow 'I don't know' and verify important facts.

Controls randomness. Low is more consistent, higher is more creative.

The maximum amount of text (tokens) the model can consider at once, including your prompt and its answer.

A piece of text the model processes, roughly three quarters of an English word on average.

Build a small test set of real inputs with expected results and score outputs whenever you change the prompt.

Text in user input or documents that tries to override your instructions. Separate data from instructions and limit what the model can do.

Retrieval augmented generation: fetching relevant documents and adding them to the prompt so answers use your data.

Give the schema, table meanings, the dialect, an example question and answer, and ask it to explain assumptions.

Say the exact length and format, for example '3 bullet points, under 50 words'.

Giving a role like 'senior SQL reviewer' sets tone and focus, but clear task and context matter more.

Passwords, keys, personal data and confidential company data unless your company approves the tool.

Change one thing at a time, test on the same examples and keep the version that scores best.