Lab: The agent brief

This lab produces the agent brief — the one document every later module amends and none contradicts. Build it for Harbor Helper using the case details, or (better, if you have one) for a support assistant your own workplace could ship.

1. 1Write the scope contract: 5 in-scope topics with a one-line 'done means' each; 6+ out-of-scope routes with destinations; the never list (aim for 8 items — pull from the lesson's list, then add two specific to the business: Harbor Lane's include 'never confirm ingredient/allergen safety — route to packaging + human' and 'never discuss the mailer-damage refund wave beyond published policy').
2. 2Write the persona spec (one page): voice adjectives with counter-examples ('warm but never chirpy: yes ‘Happy to help with that.' / no ‘Yay! Let's gooo! 🎉''), length rules, the disclosure line verbatim, and the distress-mode overrides (shorter, plainer, next-steps-first).
3. 3Hand-write 8 golden conversations covering: each in-scope topic (happy path), one clarifying-question fork, one out-of-scope redirect, one angry customer, one handoff, one 'agent doesn't know' (the honest 'I don't have that — flagging for a human' beat). Write both sides; keep them short and real.
4. 4Assemble the v0 system prompt from the three artifacts (scope → rules, persona → voice section, 2 golden conversations → few-shot examples). Run it in a chat window against 5 improvised customer messages; note where it breaks the brief. Fixing those breaks is Modules 2-4's work — today you just need the honest gap list.
5. 5Red-team the brief itself with your AI assistant: 'Here's my agent brief. List 10 customer messages that fall in the cracks — ambiguous between in and out of scope, or that exploit the never list's wording.' The cracks become eval cases and, often, brief amendments. Date and version the brief: v1.0.