Prompt Engineering Interview Pack - Medium With Hints
Level: Medium How to use: Use the hints to sharpen tradeoff language and scenario framing.
Interview Questions With Hints
Senior Engineer
- How would you design the prompt for a user asking, "Recommend something like Attack on Titan but with more politics"?
Hint: Use upstream candidate selection, a narrow recommendation-explainer prompt, and concrete similarity dimensions such as tone, pacing, and intrigue.
- How would you prompt a policy answer differently from a product-attribute answer?
Hint: Policy uses retrieved chunks and completeness handling; product attributes should prefer structured catalog fields and avoid speculation.
- What prompt changes would you make when the UI expects strict JSON output?
Hint: Explicit schema, exact top-level keys, no markdown fences, empty arrays for missing fields, and parser validation outside the prompt.
Staff Engineer
- How do you decide what belongs in the system prompt versus retrieved context versus structured business input?
Hint: Put durable rules in the system prompt, factual evidence in retrieved context, and live business facts in structured inputs.
- How would you keep a long multi-turn conversation inside token budget without losing key preferences?
Hint: Summarize history, preserve explicit preference slots, and reserve output tokens before expanding history.
- Why can adding more retrieved chunks reduce answer quality instead of improving it?
Hint: Higher recall can lower precision, increase contradiction risk, and dilute prompt focus.
Product Analyst
- What prompt-level changes would improve concision without making the chatbot feel robotic?
Hint: Set sentence targets, section limits, and priority ordering instead of only saying "be concise."
- How would you make follow-up suggestions useful instead of generic?
Hint: Tie suggestions to the actual products, policy path, or current unresolved question.
Applied Scientist
- When would you use zero-shot, one-shot, or few-shot prompting in this project?
Hint: Zero-shot for stable formatting tasks, few-shot only when it fixes a repeatable behavior gap worth the extra tokens.
- How would you adapt prompts for multilingual use without weakening business constraints?
Hint: Keep rules language-agnostic, answer in the detected language, and preserve official product or policy terms when needed for accuracy.