Grounding a portfolio chatbot so it can't be jailbroken into nonsense

How I keep an LLM that answers as me on-script: a tight knowledge brief, server-side guardrails, and refusal templates — without killing the conversational feel.