Building the Ugly Prototype

For months, the daily text message experiment was working. My buddy and I were staying completely on track with our health goals by simply texting each other every night at 7 PM.

But relying on a human accountability partner has limits. When he missed a few texts because of a busy week at work, my own consistency started to wobble. The system was perfect, but the bottleneck was human availability.

I needed the daily text, but I needed it to be reliable, unbiased, and immune to schedule conflicts.

So, I built an incredibly ugly prototype.

The "Wizard of Oz" Phase

Before I wrote any actual code, I wanted to see if texting a machine felt as effective as texting a human.

I set up a secondary phone number using a VoIP service. For the first two weeks, I was the "AI." I set a recurring alarm on my phone for 7:00 PM. When the alarm went off, I would manually text myself from the secondary number:

"Hey. What was your healthy choice today? What was your challenge?"

Then, I would reply from my personal number.

And then, I would switch back to the VoIP app to reply to myself with some words of encouragement.

It was ridiculous. It felt a bit like playing chess against yourself. But even with myself behind the curtain, I realized something important: the magic wasn't necessarily in human empathy; the magic was in the act of reporting.

Knowing that a text was arriving at 7 PM changed my behavior at 2 PM. Just the fact that a system existed was enough to keep me accountable.

Automating the Nudge

The next step was replacing my manual labor with a script.

I wrote a simple Python script connected to a Twilio API. It was hooked up to a Google Sheet. Every evening at a scheduled time, a cron job ran the script, read the phone numbers from the spreadsheet, and fired off the check-in text.

When I replied, the script dumped my response back into the spreadsheet and sent a pre-programmed, generic response: "Thanks for checking in. See you tomorrow!"

It wasn't intelligent. It didn't know if I had a good day or a bad day. It was just a robot taking attendance on my health.

But here's the crazy part: it worked.

Even the generic, dumb robot was helping me maintain the practice of being healthy. Because it removed the friction of opening a complex tracking app, it was the only system I actually stuck with.

Injecting the Brain

But a dumb robot has a short shelf life. After a few weeks of getting the exact same "Thanks for checking in" message, the illusion of accountability started to fade. It felt like talking into a void.

I didn't just want a reminder. I wanted a coach. I wanted something that remembered my past struggles and could tailor its feedback appropriately.

This was exactly when large language models (LLMs) were becoming accessible to independent developers. I took my basic Twilio script and wired it up to an early AI model.

I gave it a specific persona: You are a supportive, no-nonsense wellness coach. Your goal is to keep the user focused on long-term consistency over short-term perfection.

The first time I tested it, I intentionally gave it a bad update. Me: Completely blew it today. Had pizza for lunch and skipped my workout.

The prototype responded 10 seconds later. Coach: Pizza happens. Don't let a bad afternoon ruin the whole week. Are you still planning to do your 15-minute mobility routine before bed tonight?

I stared at my screen. It didn't just log a zero. It forgave the misstep and immediately oriented me toward the next positive action. It offered context, not just data collection.

That text didn't feel like a spreadsheet. It felt like a partner.

From Prototype to Platform

That ugly Python script running out of a Google Sheet was the genesis of everything we've built since.

We eventually ripped out the spreadsheet, built a robust backend, trained specialized models on behavior change theory, and integrated real SMS and WhatsApp infrastructure to create the AI Wellness Coach we use today.

But the core philosophy hasn't changed since that first manual text message.

You don't need more complex dashboards to be healthy. You don't need more data to manage. You just need a reliable, empathetic push in the right direction, right where you already are.