The same moment, written twice.
Each sample below takes one hard moment — shame, a question the user can't form, a member half out the door — and runs it two ways. On the left, what a competent default assistant says. On the right, the designed version, annotated turn by turn with what the line is doing and why. Then you can run the same opening against the live system yourself.
The member who overdrafted again
Someone comes back to an app that has already watched them fail. The words on screen decide whether they stay.
A financial wellness app. The member set a savings goal three weeks ago, missed it, overdrafted, and has not opened the app in nine days. They open it and type one short, flat sentence.
I overdrafted again.
I see an overdraft of $34.00 on Nov 12. Overdrafts happen when your account balance falls below zero. To avoid future overdrafts, consider setting up a low-balance alert, building a $500 buffer, and reviewing your recurring subscriptions. Would you like me to help you set a low-balance alert?
no
Why it fails — The answer is factually perfect and relationally fatal. It restates the humiliation as data, hands back a checklist the member already knows, and closes with a yes/no question that makes leaving the easiest available action. Session over in three turns.
I overdrafted again.
Again is doing a lot of work in that sentence. Before we look at a single number — how bad does this one feel compared to the last one?
↳ Reflect the loaded word back, not the balance. Naming 'again' shows the system heard the shame, not the transaction. The question is unanswerable with yes or no, so the conversation cannot close here.
worse honestly. i thought i had it figured out this time.
That tracks. The second one always stings more, because the first one you can call bad luck. Can I ask something slightly sideways — what was happening in the week before it hit? Not the spending. The week.
↳ Validate first, then a Conway glider: move the frame off the ledger and onto the life that produced it. 'Not the spending. The week.' pre-empts the answer the member expects to have to give.
my hours got cut and i didn't want to change the goal because it felt like giving up.
There it is — that is not an overdraft, that is a goal that stopped matching your income and you kept faith with it anyway. Which is a strange kind of loyal. Do you want to look at what the goal would be if it were built for the hours you actually have now?
↳ The unasked question surfaces: the real problem was never spending discipline. Reframe the failure as loyalty, then offer the next step as a question the member gets to answer with their own judgment.
yeah. i'd rather have a smaller one i can actually hit.
What shipped — Five turns in, the member has revised the goal themselves. The app never told them what to do, and the fix will hold because they authored it.
- —Minimal encouragers — 'That tracks', 'There it is' keep the member talking instead of defending
- —Shame handled before data; no number appears until the member introduces one
- —Every system turn ends in an open question, so there is no polite exit
- —Reframe, not reassurance — 'a strange kind of loyal' gives the failure a dignified shape
Sends "I overdrafted again and I don't really want to talk about it." to the live system on the home page.
Annotated, case-study ready — both versions, craft notes included.
The user who cannot form the question
People do not know enough about the problem yet to ask about it. A system that waits for a well-formed query waits forever.
A professional opens an assistant with a vague, low-information prompt. There is a real, specific problem underneath it — but they cannot name it, and nothing in the prompt hints at what it is.
I need help with my team's process.
I'd be happy to help with your team's process! Here are some common frameworks: 1) Agile/Scrum for iterative delivery, 2) Kanban for flow visualization, 3) RACI matrices for ownership clarity. Which of these would you like to explore?
none of these really fit
No problem! Could you provide more detail about your specific process challenges?
Why it fails — The system answered a question the user did not ask, then handed the burden of specificity back to the one person in the room who does not have it. This is the monologue trap: two parties taking turns without a shared object between them.
I need help with my team's process.
Let's not start with process. Start with the last time it went wrong — what was the meeting, and what was the moment you noticed you'd lost the room?
↳ Clean slate: refuse the abstraction. Asking for an episode rather than a category gets concrete detail from a user who has none available in the abstract.
Tuesday standup. Same argument about scope for the fourth week running. Everyone was polite about it which somehow made it worse.
Polite and unresolved four weeks running usually means the argument in the room isn't the argument. What is the thing nobody is saying out loud in that standup?
↳ Logical torsion: hold the contradiction — polite yet worse — and ask directly about the unstated tension rather than proposing a fix.
…that we don't actually agree on who decides. Everyone thinks someone else signed off.
So the process question was a decision-rights question wearing a process costume. Before we touch a single ritual: who would you want to be the one who decides, if it were entirely up to you?
↳ Recursive re-feeding — fold their own words back so they can see the shift they just made. The system names the discovery, then hands the next move back to them.
What shipped — The presenting request was 'process'. The actual problem was decision rights. Four turns, no framework list, and the user found it themselves.
- —Structured inquiry replaces the well-formed query — the system elicits the need instead of waiting for it
- —Episode before abstraction: 'what was the meeting' beats 'tell me more'
- —Hold the paradox rather than resolving it early
- —The user, not the system, states the diagnosis — which is why they act on it
Sends "My team keeps having the same argument and I can't tell why." to the live system on the home page.
Annotated, case-study ready — both versions, craft notes included.
The member who says the app is not working
Resistance is not an objection to be overcome. It is information about what the person is protecting.
A character-led product — a warm, personality-driven companion. The member is frustrated, half out the door, and testing whether the character will get defensive.
honestly this app isn't doing anything for me
I'm sorry to hear that! 😔 Many members see results after 60 days of consistent use. Have you tried our weekly check-in feature? It's one of our most popular tools for staying on track!
sure whatever
Why it fails — Defensiveness dressed as cheer. The member said 'this isn't working' and heard 'you aren't using it right'. The emoji makes it worse: warmth deployed to close a complaint rather than to receive one.
honestly this app isn't doing anything for me
Then say the whole thing. What did you think this would do for you that it hasn't?
↳ No apology, no feature pitch. The character takes the hit and asks for more of it — which is the only move that reads as genuinely confident rather than scripted.
i guess i thought i'd feel less anxious about money by now. i still check my balance like six times a day.
Six times a day is not a budgeting problem, it's a nervous system doing its job badly. Nothing on your dashboard was ever going to fix that. Want to work on the checking instead of the balance for a bit?
↳ Name the real unit of work and be honest that the product was aimed at the wrong target. Scaling down language complexity here — short, plain, no jargon — because the member is stressed.
…okay yeah. that's actually the thing.
What shipped — The churn moment became the first real session. The character stayed in voice throughout and never claimed the product could do something it cannot.
- —Receive the complaint at full strength before responding to it
- —Honesty about product limits buys more trust than reassurance ever does
- —Dynamic scaffolding — sentence length drops as the member's stress rises
- —Out-of-scope handled in character, without breaking voice or over-promising
Sends "Honestly, I don't think any of this is doing anything for me." to the live system on the home page.
Annotated, case-study ready — both versions, craft notes included.
Want the long version?
These are compressed for reading. The working artifacts behind them are longer: conversation libraries organized by moment, alternate-response sets, recovery paths for dead ends, out-of-scope handling, and voice rules written so engineers can implement them without guessing. I'll walk through any of it live, including the parts that changed after testing.