Mastering the Chat: Connect with Every Client and Care for Every Pet

Vet techs rarely train for chat-based client interactions, even though more clinics are moving that direction. I didn’t want this to feel like a generic branching scenario — I wanted learners to feel like they were genuinely on shift, talking to a real client about a real pet, with real stakes if they got it wrong. That meant the client’s responses couldn’t be scripted; they needed to react to whatever the learner actually typed. So I built the client conversation on GPT-4o, giving Linda a full persona and letting her respond naturally rather than branching between preset paths.

Getting there wasn’t just a matter of writing a good prompt. I needed the AI conversation to run securely and reliably inside Storyline, which meant learning to work with serverless functions — something I’d never touched before. With Claude’s help, I set up a Vercel backend to handle the API calls to ChatGPT and HeyGen, keeping the API keys secure and off the client side entirely, and used GitHub to track every change as I built it out. That backend is what made the rest of the experience possible: three connected endpoints handling the client conversation, the feedback video request, and the video status check.

One thing I cared about early on was making sure learners weren’t left to guess. In a real clinic, a vet tech unsure how to handle a situation can ask a vet for guidance — so I built that same option into the simulation. Dr. Tessa Clarke, the mentor vet, can see the learner’s full conversation with the client, not just a summary, so her advice responds to what actually happened rather than offering generic talking points. Vague encouragement like “be more empathetic” doesn’t teach anyone anything. Specific, situational guidance does.

The same thinking shaped the feedback at the end. Instead of a generic pass/fail, I wanted learners to receive feedback that reflected exactly how their specific conversation went. After the chat ends, the system generates a star rating and a review from the client’s perspective, then uses both — along with the full chat history — to write a personalized script for Morgan, the lead vet tech, delivered as a short AI-generated video. Two bugs shaped this more than I expected: a timing issue caused Morgan’s feedback video to load before the video request was fully processed, and a separate timing issue caused her feedback to read as unfairly harsh because her script was generating before the star rating had come back. Both were fixed by adjusting the sequence of events on the Storyline timeline — a good reminder that in AI-integrated projects, timing between systems matters as much as the content itself.

Once the experience was working end to end, I recorded a walkthrough and shared it with colleagues for feedback. That round surfaced a real usability gap: testers missed the notification that signaled a new message and didn’t know what to click next. I addressed it by adding a short “Meet Morgan” onboarding moment that orients learners to the interface before they’re on their own.

I also ran a live test with a small group to see how people responded in real time, and to watch how the AI video generation performed under real conditions. That session surfaced a second gap — a short activity I’d added between the client chat and the feedback video (to cover HeyGen’s processing time) wasn’t clearly framed, so testers didn’t understand its purpose. I adjusted how it’s introduced so it reads as intentional practice rather than a random detour.