Prompt Engineering Interview Pack - Basic With Hints
Level: Basic How to use: Answer out loud first, then use the hints to check whether you covered the project-specific points.
Memory Map
graph TD
A[Route First] --> B[Choose Truth Source]
B --> C[Pick Prompt Shape]
C --> D[Constrain Output]
D --> E[Validate and Recover]
Interview Questions With Hints
Hiring Manager
- What is prompt engineering trying to achieve in MangaAssist beyond "make the answer sound good"?
Hint: Cover trust, grounding, response consistency, structured output, and cost-aware use of the FM path.
- Why should MangaAssist not send every user message to a foundational model?
Hint: Mention latency, cost, hallucination risk, and the fact that many flows are better served by templates or APIs.
- What is the difference between a prompt that owns tone and a system that owns truth?
Hint: The FM can explain and summarize, but live systems and retrieved sources must own prices, order state, and policy facts.
Product Manager
- Why is recommendation prompting different from order-tracking prompting?
Hint: Recommendation needs explanation and taste alignment; order tracking needs strict factual formatting from API data.
- Why is grounding important for customer trust in a commerce chatbot?
Hint: Wrong prices, wrong policies, or wrong delivery information break trust quickly.
Senior Engineer
- Walk me through the main layers of a MangaAssist prompt from system rules to user message.
Hint: System rules, workflow context, grounding data, conversation state, current message, output contract.
- When would you use template, API plus template, RAG plus generation, and a richer FM path?
Hint: Match the answer path to the truth source and error tolerance.
- Why should business-critical facts like price or delivery status not be left to free-form generation?
Hint: Even strong prompts cannot make free-form generation a reliable source of truth for live commerce facts.
Applied Scientist
- What is the benefit of using intent-specific prompt patterns instead of one giant universal prompt?
Hint: Smaller prompts, less instruction conflict, lower token cost, and easier evaluation.
- Why can few-shot prompting improve behavior but still be a bad production choice in some flows?
Hint: Talk about token bloat, latency, context pressure, and diminishing returns.