Notes

Q. It is risky to directly use AI generated code in production without any proper input validation, error handling or handling the edge cases as it will not have a full proof of correctly working. It only full-fill happy flows. If there is no error handling and any error occurred it will crash the complete system and lead to inconsistent data processing. It may also lead to security vulnerabilities if random code without proper validationa and review is pushed to production. : A developer uses Cursor or ChatGPT to generate code for a feature. The code works in a basic test, but it has no input validation, no error handling, and no edge-case handling. Why is it risky to directly use this AI-generated code in production?** : anlayse like a strict interver and tell what point can be added

Q. A team member gives this prompt to an AI tool: “Create a good landing page.” Why is this a weak prompt? Rewrite it into a better prompt by adding clear goal, context, constraints, and expected output.

Q. An AI assistant summarizes a client document but adds details that were not present in the document. What is this problem called, why is it risky, and how can the team reduce this risk?** : answer in 4-5 lines

Q. Explain the difference between a system prompt and a user prompt. Give one simple example of each for an internal tech-support AI assistant.**

Q. Your company wants an AI assistant to answer questions from internal policy documents, not from general internet knowledge. What approach should be used, and what instructions should be given to the AI to reduce wrong answers?

Q. An AI workflow is expected to return customer details in a fixed JSON format, but sometimes the AI returns plain text instead. Why is this a problem for automation, and how would you make the output more reliable?**

Q. An AI agent can read customer emails and perform actions like sending replies, updating booking status, or forwarding messages. What guardrails should be added before allowing the agent to take actions automatically?**

Your team has built an AI feature that answers user questions. Before releasing it, what types of tests would you run to check whether the AI feature is reliable, safe, and useful?**


Q. You are using Cursor or any AI coding assistant to build a backend API endpoint.*

*Requirement:*
Create an API endpoint that receives a customer message and returns one of these categories:
"bookingquery", "cancellation", "paymentissue", or "other".

*Write a clear prompt you would give to the AI coding assistant so that it generates reliable code.*

*Your prompt should include:*
- the goal
- expected input
- expected output
- error handling
- simple validation rules*


Q. An AI model is expected to return this JSON:*

{
"category": "payment_issue",
"summary": "Customer is asking about failed payment.",
"confidence": 0.86
}

*Write simple pseudo-code or code to check that:*
- category is present
- summary is present
- confidence is present
- confidence is a number between 0 and 1

*If any check fails, the workflow should send the response for human review.**


Q. An AI model classifies a customer message and returns:*

{
"category": "cancellation",
"confidence": 0.72
}

*Write simple pseudo-code or code for this routing logic:*

- If confidence is 0.85 or above, continue the automation.
- If confidence is between 0.60 and 0.84, send it for human review.
- If confidence is below 0.60, ask the user for clarification.*


Q. An AI coding tool generated code for a login form. The code works in a basic test, but it does not validate empty fields, does not handle API errors, and shows the same message for every failure.*

*Write the improvements you would ask the AI tool to make before this code is production-ready.*

*Your answer should mention:*
- input validation
- error handling
- user-friendly messages
- security or reliability considerations*


Q. An AI agent reads customer emails and can perform actions like:*
- send a reply
- update a booking status
- forward the email to another team

*Write simple pseudo-code or workflow logic for this agent.*

*The logic should ensure:*
- low-risk replies can be sent automatically
- payment, cancellation, or refund-related cases go to human review
- the agent does not update booking status unless confidence is high
- every action is logged*