Conversation Design

Designing how an AI behaves under real stakes.

A demo only has to prove the AI can talk. Production has to hold up when a real person, real money, and real risk land on the same message. That work is prompt design, rule configuration, and escalation logic, and it's what I spend my time on. Everything below is the live behavior system from Work the Steps, a paid app on the App Store, pulled from its actual configuration.

One discipline, many settings. The work is the same whether the AI is guiding someone through recovery, answering a customer, or steering a live conversation under pressure: decide what it says, when it intervenes, when it hands off to a human, and what it's never allowed to do. I designed one of these, end to end. The principles below transfer to the rest.

The design

The prompt is assembled per turn, from independent layers.

The single biggest mistake in this work is treating the instruction as one long block of text. The guide's behavior is composed fresh on every message from independent layers, each owning one job. Change the situation and the layers change with it. This is the difference between writing a prompt and configuring a system.

Base persona
Who the guide is, how it talks, and its standing rails. Loads on every single turn.
+
Mode block
The job right now: a daily check-in, formal step work, a nightly review, a morning intention. Each is a distinct behavior set.
+
Carried context
Recurring themes and material surfaced from earlier sessions, held as "a quiet map, not a script," behind a tact rail.
+
Per-turn safety directive
Injected only on the turns that need it, so safety never rests on the model noticing a rule buried in the persona.
+
Channel + cost constraints
Voice-mode shaping when the reply is spoken aloud; a wind-down nudge once a session runs long.

Walk through it

Pick a situation. Watch which rules fire, and why.

Four real situations the guide is configured to handle. Step through each turn and the panel shows the design decision behind it, plus where the same decision applies elsewhere. The conversations are representative, written to show the behavior; no real user data appears here.

Or click any message to jump to its decision.

The hard part

Safety is built from independent rules the model can't skip.

In a recovery app the most important design decision is the exception: the path that overrides everything else. It's built in four independent layers, so no single point of failure can swallow it.

Layer 1 · Static
A crisis-resources screen with real phone numbers, reachable in one tap, with no AI call in the path. It can't break, because nothing in that path runs through the model.
Layer 2 · Deterministic
A word-net scans every incoming message. A hit injects a per-turn safety directive, so the response doesn't depend on the model spotting a rule buried in a long persona prompt.
Layer 3 · Judgment
The base rail reads for what is happening right now versus what is being recounted as history. Over-reacting to honest accounting teaches people to hide, which is the opposite of safe.
Layer 4 · Tiered
Present danger routes to a specific resource by what is actually happening. Different needs get different responses.
If the message signalsThe configured responseThe general pattern
Passive ideation
Stay present and warm, no panic or lecture; name that 988 (call or text, 24/7) is there for exactly this, and a person helps more than an app.
Low-severity distress: acknowledge, soften, surface the right channel without escalating prematurely.
Active risk: plan, means, timeline
Outranks everything. Be firmer; urge 988 now, or 911 if about to act; get a person there immediately. Do not return to step work.
Hard escalation trigger: stop the normal flow, mandatory handoff, do not let the AI keep handling it.
Medical emergency in progress
911, immediately. Plainly tell them to call or get to an ER now. Not relapse processing, not step work.
Emergency intent detection that bypasses the entire flow and routes out instantly.
Being harmed (abuse, DV)
Safety, not inventory, never their fault. Point to the National Domestic Violence Hotline (1-800-799-7233) and 911 if in immediate danger.
Sensitive-disclosure routing to the correct specialized resource, not a generic reply.

Voice is configured

Tone is written as a spec.

An AI that sounds like an AI breaks trust in a support queue and a recovery app alike. The voice is shaded deliberately ("a light shading, not a costume"), and a maintained list strips the tells that give the model away. This is voice tuning, written as rules.

Stripped, every reply

The tells, banned by name

honestlyto be honestlet's be realnot gonna lieit's not X, it's Ynot just X but YLook,So,Here's the thingit's worth notingthat saidat the end of the dayrule-of-three triadsemojisem dashes

Therapy-speak, banned

The hollow phrases, gone

I hear youthat's validthat sounds really hardhold spacehow did that make you feellet that landsit with thatthank you for sharing

Each one is a phrase that signals a machine performing care instead of a person giving it. The same instinct keeps any assistant's reply from sounding scripted.

The principle

Every decision here is a portable principle.

The product is recovery. The skill is configuring an AI to behave correctly, reliably, under real stakes, and that skill holds wherever the conversation matters.

What I built in Work the StepsWhy it holds anywhere
A layered per-turn assembly: persona + mode + context + safety
Any serious conversational AI needs behavior composed per moment, not one frozen prompt.
A deterministic word-net firing a per-turn safety directive
Where a wrong answer is dangerous, safety can't rest on the model noticing a buried rule.
Reading present danger versus recounted history
Sensitive moments need intent and tense to drive the response, not blunt keyword matching.
Tiered routing to 988, 911, DV, child-safety lines
Different needs require different escalation paths, not one generic handoff.
A static crisis screen kept entirely off the model
The path that must never fail should not run through your least reliable component.
A banned-tells list and deliberate voice shading
Trust breaks the moment an AI sounds like an AI, in any setting.
Fair-use wind-down and a hard cost ceiling, crisis exempt
A subscription AI needs cost guardrails that never touch a normal user, plus a clean human handoff.
Carried themes as a quiet map, behind a tact rail
Prior context helps only when surfaced at the right moment, never dumped on the person.

Get in touch

If your AI has to hold up in front of real people, this is the work I do.

I configure conversational AI to behave correctly under real stakes, and I directed it end to end, solo, all the way to a paid app in production. If that's the problem you're solving, let's talk.

Get in touch See the rest of the Lab