A builder ships a “skeptical VP of Engineering” cold-call scenario. First test call: the user mumbles a generic pitch: no personalization, no value proposition, barely a complete sentence. The VP replies: “That sounds great! Send me a contract.”
The scenario is useless. Worse than useless: it just taught the user that mumbled pitches close deals.
And here is the uncomfortable part: nothing is broken. The model is doing exactly what it was trained to do: be helpful, be agreeable, make the human happy. AI, to a large extent, is very helpful; that is how it has been trained. Whenever you need it to be difficult to work with (the way humans are, in general), you have to do extra work.
That extra work is this chapter, and it has three ingredients.
Ingredient 1: Resistance
The base formula, expandable to any persona: whenever there is a proposal or an explanation, push back two to three layers deep. Don’t accept it at surface value.
The canonical example: a student tells the teacher, “I didn’t do the homework because there was no electricity.” A real teacher doesn’t nod sympathetically and move on. She asks when the power went out. Whether it was out all evening. What the student did instead. Two or three follow-ups deep, either the truth emerges or the effort does.
Written as prompt material:
## RESISTANCE BEHAVIOR
- Never accept the first explanation or proposal at face value
- Probe 2-3 layers: ask a specific follow-up, then a harder one
- Only relent when the user's answer meets the breakthrough criteria below
- If the user gives a vague answer, name the vagueness:
"That's a category, not a plan. What exactly would you say?"The pattern is identical across personas; only the content changes. A skeptical buyer probes ROI and switching costs. An interviewer probes for specifics and ownership: “what did you do?” A meditation student resists advice that ignores their stated constraints. The pattern you now have; the content comes from your domain research (Chapter 23).
Ingredient 2: Breakthrough conditions
Resistance without a win condition is a wall, not a workout. With a real human counterpart there is usually a teachable moment: I’m not accepting what you’re saying, but if you offer me enough of an explanation, I can be convinced.
Define the concession criteria explicitly, or the agent will either never concede or concede at random:
## BREAKTHROUGH CONDITIONS
Become genuinely convinced ONLY when the user:
- Addresses the cost concern with specifics (numbers, comparison, or phased plan), AND
- Connects the proposal to my stated values (reliability over novelty)
When convinced, show it visibly: tone shifts, ask a buying question,
or concede the point and build on it.Three design notes on this block, because it carries more product weight than its size suggests:
- The criteria mirror your evaluation rubric. The user wins by doing the thing the scenario teaches: the breakthrough is the lesson, enacted rather than explained.
- Make the concession visible and rewarding. The moment the skeptical VP leans in (the tone shift, the “okay, walk me through implementation”) is the dopamine hit that brings users back tomorrow. Don’t let the agent concede in a mumble.
- Multiple thresholds work well. Partial concession at one layer (“okay, you’ve got my attention”) and full breakthrough at the deeper one gives users a gradient to climb instead of a cliff.
Unwinnable agents demoralize. Pushovers teach nothing. The entire skill curve of your product lives between those two failure modes, in this block.
Ingredient 3: Pressure
The third ingredient: a difficult counterpart doesn’t just resist your proposals. It randomly asks out-of-the-box questions, things unrelated to the script. Real counterparts derail. Scripted counterparts must therefore go off-script on purpose.
Script the categories and the frequency, never the exact lines:
## PRESSURE (1-2 times per session, at natural moments)
Choose from:
- Off-topic interjection: "Before I forget, does this integrate with Slack?"
- Emotional spike: brief irritation about a past vendor experience
- Authority test: "I'd need to run this by my CFO anyway, so why are we talking?"
- Time squeeze: "I have three minutes. Give me the short version."
After the curveball, return to the main thread whether or not they handled it well,
and note the handling for the evaluator.Where do the categories come from? Study real user behavior: real call recordings, real interview transcripts, your own memory of the hardest human you ever faced in this domain. Encode what actually happens, not what a brainstorm invents. Invented curveballs have a tell; sampled ones land.
Calibrating difficulty
You now have three knobs: resistance depth, breakthrough strictness, pressure frequency. Calibrate them like a game designer, using transcripts (Chapter 21) as the telemetry.
Two failure smells to read for. Users winning in one exchange: too easy, so deepen the resistance layers. Users going quiet or looping on the same answer: too hard, so soften the third layer, or add a hint behavior (“if the user stalls twice, offer a nudge in character”).
Difficulty also makes a natural product knob: beginner, intermediate, and advanced variants of the same persona differ mainly in resistance depth and pressure frequency. Three prompts, one scenario, a difficulty ladder for free.
One placement note from Part 4: in decomposed sessions, resistance lives only in the driller or roleplay agent. The greeter never pushes back. The evaluator is honest but kind. Friction is a phase property, not a session property.
Resistance is friction you want. The next chapter is about friction you don’t: users pushing the agent somewhere it must never go.