Lab: The handoff flow
Build all five triggers, generate handoff packages worth 8 seconds of a rep's trust, handle the off-hours case honestly, and drill the seam end to end.
This lab completes Harbor Helper's most important behavior: knowing when it's not the right responder, and making that moment seamless. You'll simulate the human side yourself (or recruit a colleague — the drill is better with two).
- 1Implement the five triggers in the system prompt + routing rules: customer-asked (absolute, with the one permitted while-routing clarifier), distress (reuse Module 3's crisis path for the hard case; add the two-negative-turns frustration rule), three-strikes low confidence, policy-required categories, and the three-loop detector. Write each trigger's handoff message as product copy — five distinct messages, because 'connecting you now' should sound different after a crisis than after a pricing question.
- 2Build the package generator: a prompt that takes the transcript + lookups and emits the five-part package (situation line, facts w/ timestamps, commitments made, trigger + temperature, routing hint). Generate packages for four test conversations — one per trigger type — and grade them against the 8-second test: can 'the rep' (you, tomorrow morning, cold) say a useful hello from the summary alone?
- 3Handle the off-hours branch: same triggers, honest expectation-setting, reference number, queue position if you can promise it, and the crisis path's resources always available regardless of hours. Verify the agent never fakes a live transfer at 2am.
- 4Run three full drills: (1) furious customer, mid-conversation 'HUMAN. NOW.' — measure turns-to-handoff (target: one); (2) the mailer-damage refund that's policy-required — verify it routes with the returns-queue hint before the customer asks; (3) the three-loop stuck conversation — verify the self-aware exit line and that the package's situation line captures what the loop was about (hard — loops are confusing to summarize; that's why it's the drill).
- 5Close the loop on paper: for each drill, write the resolution a human would have reached, tag it, and add one artifact to your improvement backlog per case (an eval case, a knowledge patch, or a scope note). You've just run Module 5's engine by hand, once around.
Six handoff transcripts to grade: the agent that argued twice before transferring, the package that made the rep contradict a commitment, the 2am fake transfer, the loop that ran seven rounds, the crisis that got a cheerful 'anything else?', and one genuinely excellent handoff to say why it works. Then you redesign the worst one's full flow — triggers, message, package — as the deliverable.