Two versions of the same coaching session.
Session A opens: “Hello! How can I help you today?” The user flounders: help with what, exactly? The middle is twenty minutes of undifferentiated talk. It ends when the user awkwardly says, “um, I guess that’s it?”
Session B is Chapter 1’s demo: greeted by name, agenda set in one sentence, a picture painted, an option card offered, pressure applied at layer two, and then: “You’ve got a solid escalation ladder now. Next time, parent communication. See you then.”
Same model. Same tools. Same minutes. The difference is that Session B’s flow was designed and Session A’s was hoped for. The design directive has three parts: define how the session opens, what happens during it (the objection handling, the breakthroughs), and what happens at the end. And at every transition, there should be some change in behavior. This chapter takes the three parts in order.
Openings: the first 60 seconds
Three jobs, in order, all inside a minute:
1. Continuity. Use super memory, by name and by thread: “Last time we worked on X, and you wanted to try Y.” Chapter 11’s re-onboarding anti-pattern (the returning user answering the same questions forever) dies in this sentence or nowhere.
2. Agenda. One sentence on today’s shape: “Today: one new scenario, then we’ll drill the part you found hard.” Users relax when they know the container. Anxiety about what’s coming is attention stolen from the work.
3. The first easy turn. End the opening with a question the user cannot fail: “Ready?” or “Pick up where we left off, or try something new?” Momentum before difficulty: resistance belongs later, and per Chapter 18’s placement rule, the greeter never pushes back.
First-ever sessions have no memory to draw on, so swap continuity for a 30-second orientation: who the agent is, how the session works, and what to do when stuck. Then the same easy first turn.
Middles: alternate the modality
Voice-only middles sag. It is Chapter 13’s attention problem, but on the human side: twenty minutes of unbroken talking blurs for the listener just as surely as an unbroken transcript blurs for the model.
The fix is the multimodal toolkit, deployed rhythmically:
- Talk: present, explain, discuss. Keep each unbroken stretch under about two minutes.
- Show: the image tool paints the scene at every new scenario. Chapter 8’s frequency instruction did this on purpose: “at least once per scenario” is a rhythm rule wearing a tool instruction’s clothes.
- Check: an MCQ pins a decision point, and the options stay on screen while the user thinks. Then the demo’s follow-through rule: whatever they pick, ask them to justify it.
- Do: the roleplay or drill, where Chapter 18’s resistance and pressure live. The user speaks as themselves inside the scenario, never summarizing. “I would say something about value” is not practice; make them actually say it.
Order matters: comprehension before performance. Show and check before you make them do: users who fail a drill they never understood learn only frustration.
And mark every transition audibly: “Now let’s switch from theory to practice. I’m the customer now.” The acceptance test for your whole flow design: there should be some change in behavior at every seam. If a transcript reads the same in phase two and phase four, your flow exists in the prompt but not in the experience, and the prompt is decoration.
Endings: land the session
Sessions must end; they must not evaporate. The deliberate close has four moves, in order:
1. Summarize the work. Two or three specifics, never generalities: “Your acknowledge-and-reframe is solid; discounting under pressure is the pattern to watch.”
2. Preview the next session. Continuity, pointed forward: “Next time: parent communication.” Notice this sentence does double duty: it also writes itself into memory as Chapter 11’s what-to-cover-next field. The ending of this session is the opening of the next one, drafted early.
3. Write memory. The agent’s exit action per Chapter 12: observations, progress, preferences, steered by your memory instructions.
4. End deliberately. Invoke the end-session tool after the goodbye. Chapter 8’s frequency line was written for this: exactly once, only after the wrap-up summary. A session that just stops mid-air feels broken even when nothing broke.
One timing rule protects all four: the wrap-up starts before time runs out. Give it a protected 90-second budget in the flow. A rushed close undoes a great middle: the last thirty seconds are what the user walks away holding.
Example: the full annotated flow for Chapter 4’s scorecard winner, cold-call practice for SDRs: Opening (memory: “last time you stalled on the gatekeeper”) → Show (image: the prospect’s office, mid-Monday chaos) → Check (MCQ: pick your opener) → Do (dial-tone roleplay, VP persona, pressure at minute four) → Breakthrough or stall → Close (“two wins, one pattern to fix. Next time: voicemail scripts”). Every arrow is a change in behavior. Every element traces to a chapter you have now read.
You now have the full design vocabulary: workspace, memory, sub-agents, friction, flow. Part 6 is where designs meet reality: the iteration loop, the tools that earn their keep, and the domain research that fills the one gap no pattern can.