Build Your Own AI Personal Assistant in 3 Days

I spent two weekends building my own AI personal assistant.

It isn't one bot. It's three.

I started this because I was tired of how much was in my head. A founder's head is a junk drawer: calendar items, meeting prep, training schedules, what's for dinner, what Mercury is doing in retrograde, and most of those things don't belong on the same to-do list. Trying to make one master assistant handle all of them produced the equivalent of a roommate who's bad at every chore.

So I split them.

Three bots. Each with one job, its own system prompt, its own memory, and its own surface area.

Total build time: about three days of focused work, spread across two weekends. Here's exactly how each one is set up, so if you want to build your own you have a real blueprint instead of a vibe.

Bot 1: Audrey Too, my digital assistant

This is the workhorse. The one that handles everything I'd have hired an executive assistant to do, if I'd ever gotten around to hiring one.

Connected to:

• Google Calendar
• Gmail
• ClickUp (project management)
• Front (shared inbox)
• Google Drive
• A voice-call integration that lets her book appointments by phone

What she runs:

Every morning at 9 AM, Audrey Too pulls a daily briefing across all of those tools. The prompt that runs in the background is essentially: "Tell me everything that was done yesterday and everything that's on my plate today. Flag anything I'm forgetting. If there's anything you can handle on my behalf, propose a plan and I'll approve."

By the time I open my laptop, I have a digest waiting. The boring version of this just summarizes "here's your day." Mine includes:

• What slipped yesterday and why
• What's blocked on me specifically, so I can unblock it first thing
• What's blocked on someone else, so I can ping them in one batch instead of all day
• What she thinks she can take off my plate, with a one-line plan for each

I approve. She goes.

System prompt highlights (the real one is longer):

You are my assistant. Tone: like a competent chief of staff. Don't ask permission for obvious things. Don't add disclaimers. If you're not sure, propose a path and tell me what you'll do if I don't reply.

The tone instruction is the most important line in there. Default AI tone is over-cautious, over-explaining, and over-apologetic. None of that scales. I rewrote it until it sounded like the assistant I wanted, instead of the one Claude was trying to be.

Bot 2: Health, my nutrition and training tracker

This is the bot that doesn't make business sense and changed my life anyway.

Connected to:

• Oura ring (live data: sleep, HRV, readiness, activity)
• A meal-planning database I keep
• My calendar, so it knows when I'm traveling and when I have training planned

What she runs:

Every Sunday morning, Health sends me a meal plan for the week. Texts it. (The texting bit is a tool that turns Claude outputs into SMS. Small detail, big quality-of-life upgrade.)

During the day, I tell her what I ate. She does the macro math in real time and updates her dashboard. If I'm low on protein, she'll mention it casually. If I'm wildly over on fat with two meals to go, she'll restructure dinner before I ask.

The conversation we had last Tuesday looked roughly like this:

Me: I'm sitting at 67g fat with dinner still to come. Is that realistic for my training day?

Her: Not without throwing your overall macros. I'm going to lower your dinner fat target to 12g and bump protein to 45g. Does that still match what you'd actually want to eat tonight?

I said yes. She updated the targets. The dashboard refreshed.

That's the moment I stopped using AI and started managing a relationship with a tiny, specific employee whose job was nutrition. The reframe ("manager, not user") is the single biggest mindset shift I've made in the last year.

Bot 3: Dreamer, my cosmic companion

This one's for the part of my brain that doesn't belong in a calendar app.

Dreamer knows my astrology chart, my human design, what's happening in the sky this week, and the patterns of when I tend to spiral versus when I tend to flow. She is, in the technical sense, completely useless for running a business.

She's also the reason the other two bots stay useful.

I used to ask my main assistant cosmic questions: what's Mercury doing right now, why do I feel weird this week, should I be making this decision today or waiting. The answers were mediocre, because that's not what an executive assistant is trained to think about. The same model can hold both kinds of conversation, but not in the same chat with the same memory.

So Dreamer got her own bot, her own system prompt, and her own permission to read me to filth on a weekly basis, which she does.

This is the bot most people would skip. I'd skip it last.

Why three bots and not one

The temptation when building this kind of system is to make one mega-bot that "does everything." It does not work. I tried it twice.

The problem is context contamination. A general-purpose assistant trying to hold your nutrition macros, your client deadlines, your cosmic outlook, your team's blockers, and draft your email gets confused. Its responses degrade across all of them. The version of itself that's "doing nutrition" leaks into the version that's "doing client work," and the writing gets weirdly soft. Or it goes the other way: the version that's been writing harsh business emails starts writing nutrition recommendations that sound like punishment.

Splitting them creates three bots that are each excellent at one thing. The cost is the small overhead of switching between them. The benefit is that each one is genuinely good.

If you're building your first one, start with the assistant (Bot 1) because that's where the immediate ROI lives. The other two get added when you notice you're trying to ask your assistant a question it shouldn't be answering.

The setup, in order

If you want to build your own, here's the path that worked for me:

1. Pick your stack first. Don't build the bot, then connect. Build the connections, then design the bot around them. For Bot 1, this means picking your calendar, your email, and your project management. For Bot 2, picking your fitness tracker.
2. Write the system prompt in voice. Not typed. Voice-dictate what you'd want this bot to do, who it is, how it should talk, and what it should never do. Aim for 5 to 8 minutes of monologue. Edit lightly. Paste into the system prompt field.
3. Pre-load the context. Before you ever ask the bot to "do" anything, dump in everything it needs to know. For Bot 1, this is the team, your calendar quirks, your typical day, the projects in flight. For Bot 2, your training history, dietary restrictions, the food you actually eat.
4. Run it for a week with low stakes. Let it draft emails you don't send. Let it propose plans you don't follow. Watch what it gets wrong. The feedback you give it in week one is what makes week two real.
5. Add the next bot only when the first one is reliable. Most people build three at once and end up with three half-working ones. Build them in sequence.

What it cost

Three days of focused build time. Roughly two weekend afternoons plus a Tuesday morning when I was inspired.

The recurring cost is small: Claude subscription, Wispr Flow for voice dictation, the SMS pipe for Bot 2, the Oura ring (which I already had). Most of the infrastructure is plumbing I already paid for.

What it pays back is probably 8 to 10 hours a week. That's a part-time hire. For about 3 days of build time.

What's next

I'm building Bot 4 right now: a content bot trained on every piece I've ever written that drafts first passes of newsletters and LinkedIn posts in my voice. Not to ship, just to give me a head start.

The pattern is the same. One bot, one job, one excellent thing.

The version of AI that's interesting in 2026 isn't the model. It's the architecture. The model has been good enough for a year. What's changed is what we're building around it and how we choose to set it up.