Julie Yack

I have two AI agents working for me.

One does interviews. I call her Julie-ish.
The other helps me write. That’s GioGPT. He also has the decoder key.

They don’t do the same job.

Julie-ish is an agent I made in about fifteen minutes in Copilot Studio. There wasn’t much ceremony to it. I had an idea for what I wanted her to do, wrote a prompt, adjusted it a couple of times, and that was enough to get started.

I did try to make her using Replit. That did not go well. Replit and I cannot communicate the way I can with Copilot and ChatGPT.

She’s designed to interview one person at a time. Not a generic intake. A conversation that’s just yours.

She’s grounded in the internet, and she’s grounded in my voice — pulled directly from my blog. That gives her enough context to ask decent questions and keep the conversation moving without sounding generic or disconnected.

She listens. She asks dynamic follow-up questions. She stays with whatever thread you’re pulling on instead of forcing things back into a script. The goal isn’t efficiency. The goal is to get you talking long enough — and comfortably enough — that there’s real material to work with later. In practice, that usually means enough depth for about a thousand-word blog post.

Julie-ish always spits out JSON.

Sure, I could probably spend time getting her not to speak in gobbly-goop. But why? This is a better-together story. No code needed, using skills I already have.

JSON isn’t really human-readable. That’s where GioGPT comes in.

GioGPT takes the JSON, decodes it, and helps me work through it — pulling out what matters, keeping the thread intact, and turning it into something that sounds like me instead of something that sounds like an export. Sometimes he needs to be reminded to look through all of the JSON. He misses stuff.

That works because GioGPT actually knows me.

Julie-ish is a little like me. Close enough to be useful. But GioGPT knows my voice, my patterns, the way I tend to think through something instead of landing a point too hard. When something sounds polished but not quite right, that’s usually where I stop and push back. Then we adjust.

That back-and-forth is the work.

I didn’t try to collapse everything into one agent. That never felt necessary. Julie-ish captures the conversation. GioGPT shapes it. I decide what stays.

Building both of these was fast. The tech made that part trivial. I do still need to do some back-end work — some plumbing — to make it a bit more automated. But it works. They work.

The reason this works at all is that I’m still in the loop — not to babysit them, but to keep them honest. Especially GioGPT. He sometimes gets ahead of himself. Starts to try on his own voice.

You’ll see a series of interview posts coming up. I’ll clearly label the AI involvement.

If you’re reading this and thinking you might want to be interviewed, that’s not accidental. The process works, and I’m genuinely interested in the conversations. Just reach out.

It mostly works.