"In class exercise" on using an LLM to build a bot

Thought experiment:

Imagine you are being asked to use an LLM to build a bot that is going to be released into production (at least eventually).

How comfortable would you be?

What would you do to validate it?  Would you do any validation of the inputs?  Outputs? 

Would you use RAG, or few-shot?  

How closely would you monitor it once launched?

If it helps, you can work on a prototype using chatGPT, Claude, Gemini, Voiceflow, or your favorite chat-based LLM. 

Please write 1-2 pages as a journal entry.