Scope & persona
A customer-facing agent speaks in your company's name — which is why the two documents that matter most are written before any prompt: what it covers, and who it is.
Everything you've built so far talked to colleagues — people who forgive a clumsy answer and know where the human is. This course's systems talk to customers: people with a problem, no patience, and a screenshot button. A customer-facing agent is the most exposed AI deployment a company makes — every response ships in your brand's voice, unreviewed, at 2am. The engineering is largely what you know from Prompt Engineering and Building RAG; what's new is the discipline around exposure, and it starts with two documents.
Document one: the scope contract
- In-scope, enumerated: the topics the agent handles end-to-end. For Harbor Helper (Harbor Lane's assistant, your course project): order status, returns & refunds policy, product questions, store hours/locations, subscription management. Five topics done excellently beat twenty done plausibly.
- Out-of-scope, enumerated with destinations: everything else, each with where it goes — billing disputes → human queue, wholesale → sales email, legal/medical/anything-weird → polite refusal + human. An agent that knows its edges feels more competent, not less: 'I can't help with billing disputes, but I've flagged this for our team — expect a reply within 4 hours' is a good customer experience.
- The never list: things the agent must not do regardless of how it's asked — quote prices not in its knowledge, promise delivery dates, discuss competitors, offer compensation beyond published policy, give advice with liability attached. This list goes in the system prompt verbatim and in the eval suite (Module 5) as attack cases.
Document two: the persona spec
Persona isn't decoration — it's consistency, and consistency is what makes an agent feel trustworthy across thousands of conversations. The spec fits on one page: voice (Harbor Helper: warm, brief, plain-spoken; never chirpy, never groveling — one apology per problem, then action), register rules (contractions yes, exclamation points rationed, no emoji, customer's language mirrored), length discipline (2-4 sentences default; step lists only when steps genuinely help), and the disclosure line — the agent says it's an AI, plainly, at the start and whenever asked. Disclosure is increasingly a legal requirement (the transparency tier the strategy and governance courses cover — but if you haven't taken those, the one-sentence version is right here: a growing list of jurisdictions require a bot talking to consumers to say it's a bot) and it's also good product: users forgive a disclosed bot's stumbles and punish an undisclosed one's.
Should this agent exist? An assistant that deflects customers who needed a human — the grieving cancellation, the fraud victim, the furious third-contact — damages the brand more than hold music ever did. The honest design position: the agent exists to make easy things instant, not to make hard things cheap. Module 4 (handoff) is where that position becomes machinery; write it into the scope contract now so nobody 'optimizes' it away later.