What does your assistant say when someone has made a mistake, or feels ashamed? Knowing the human side — humor, timing, the in-between spaces — is what makes a system work. I design the dialogue that engages people when they feel dismissed, disinterested, low, or stuck.
This is the dialogue protocol running live: a clean slate, emotional sensing, one unexpected Socratic question at a time, and a sideways "glider" prompt when you stall. Bring something genuinely unresolved and watch it work the unasked question out of you.
See annotated dialogue samples →Nothing is stored. Not a substitute for professional care.
I found this while building a human-centered dialogue system, adapting a colleague's research. The idea is simple: people don't know what they don't know. Sitting in front of a chat window, they don't know what to ask. They can't type the question. So the dialogue goes dark and they abandon it.
The usual failure looks like this — a person types something small, gets a generic answer back, doesn't feel understood, and walks away. No amount of model quality fixes a badly formed question. The fix is in the writing: the system has to draw them back in with curiosity, with a question that intrigues, with the conversational equivalent of a raised hand.
"Any teacher knows that silence means you've lost the class. You want those hands to go up. Interaction equals learning — and a chatbot is no different."
Each one was built around a specific hard moment. Sample transcripts are being assembled; ask and I'll walk you through any of them live.
A thinking partner rather than an answer machine: emotional sensing, one unexpected Socratic question at a time, and a sideways 'glider' prompt when the user stalls. Running live on this site.
What does dialogue do when the person doesn't know what to ask?
Try it →A guided intake becomes a written narrative, the model then interrogates that narrative from several angles, and the result breaks into a ladder of small steps and a vision board. A real privacy fork up front: stealth for the tab only, or save on this device. No account, no email, no database.
What does a system do when the person's own story is the material — and the AI's job is to question it, not to write it?
Ben Smith's vision, adapted and built by Lee Chazen (with support from Claude AI, Claude Code, Gemini Notebook and Lovable AI).
Try it →Built with a PhD psychotherapist to train people on the EFT tapping technique, now rebuilt on the Tessera protocol: somatic holding space, a Socratic reframe before any script, a live meridian map, and 0–10 check-ins folded back into each round.
What does the character say when someone is ashamed, resistant, or about to quit?
Try it →A future-self dialogue companion designed to make long-term thinking feel personal and immediate. Uses voice, constraint, and reflection to turn goals into conversation.
How do you get someone to listen to a version of themselves that does not exist yet?
Try it ↗A study and retrieval coach built around spaced repetition, elaboration, and the emotional moment when a learner realizes they do not actually understand something.
What does the tutor say when the student has been fooling themselves?
Try it ↗A tool that analyzes the slant of an article. The design problem was defensiveness: the tone had to stay neutral and inquisitive while challenging a source the reader trusts.
How do you question someone's belief without them closing the tab?
Try it ↗A geopolitical learning game built in under five weeks, generating 400+ curriculum-aligned scenarios on a recursive dialogue system. Average session: about thirty minutes. Falling Walls finalist, top 70 of 1,000+ entries.
Can a conversation hold attention the way a game does?
Try it ↗Both my parents worked in mental health; my own training was in education psychology. That's where I learned techniques like minimal encouragers — the small moves that keep a conversation going. Seven years of teaching taught me the rest: when the lecture failed, I drew students back in with a game.
With a PhD therapist I built a tool that trains people on EFT tapping. Clinical transcripts became multi-turn conversations for anxiety and resistance, with real safety boundaries. Twenty hours of budget, and a character that people actually trusted.
A decade inside cash management, interest rates, the Federal Reserve, FINRA and SEC guidelines. The same principle as a classroom: people want to learn things in a language that makes sense to them, on their terms. Every sentence warm, plain, and compliant.
Two years studying co-creative intelligence — a delicate give and take that feels more like jazz than a query. Decades of music practice finally pay off here: AI understands metaphor like nobody's business, and a musical frame gives a system more than description alone can.
Contract, retainer, or a single evaluation — all of it is open. Rates below are indicative starting points; scope and terms get settled in correspondence after you write. Everything begins with the same form.
Fifteen minutes. Where is your assistant getting stuck, what are its voice guidelines, and is this a fit? No deliverable, no obligation.
A focused read of one conversational system — the dead ends, the generic responses, the moments a person feels unseen — plus a written paper on what to change and why.
Working sessions with your team on elicitation, minimal encouragers, and the design of hard moments. Half day, full day, or a recurring cadence.
The actual writing: persona, voice guidelines, multi-turn flows, safety boundaries, and the recovery paths for users who are confused, discouraged, or embarrassed.
A standing block of hours each month for iteration, transcript review, and advisory — or a longer contract role. Both are open right now.
Not sure which of these fits? Skip all of it. Describe what you're working on — where your assistant's voice gets stuck — and I'll read it myself and tell you honestly whether I can help.
Shipped applications, written work, and what the people who've worked with me actually say.