Building Your Own AI to Run Your Work

If you want to build your own personal AI to run your work, the move that actually matters is this. Stop handing it tasks one prompt at a time, and start building it into tools that know your goals, your projects, and your history. A chat window forgets you the second you close it. A system you build remembers. That gap, the gap between renting a smart stranger for five minutes and slowly growing something that knows the shape of your whole life, is the entire game. I’ve spent the better part of a year building one, and I want to walk you through what it is, what it took, and where it went wrong, because that last part is the part nobody tells you about.

Why build tools instead of just chatting

The chat window is where most people stop, and I understand why, it’s incredible on its own. But I noticed something after a few months of living in it. Every conversation started from zero. I’d explain my projects again, my goals again, what I was in the middle of again, and by the end I had spent more energy briefing the thing than I saved by using it. Hell, some days the briefing was the work. And that’s the tell. The thing that turns AI into an amplifier isn’t the model being smart, the model is already smart, that part is solved. It’s the model knowing your context without you re-explaining it every single time.

So the real work isn’t learning to prompt better. It’s building the scaffolding that holds your context, so the AI walks into every session already knowing who you are and what you’re trying to do. When I started thinking about it that way, the whole project changed. I stopped trying to write the perfect prompt and started building a thing that wouldn’t need one.

There’s a quieter reason too. I’ve worked with data for well over a decade, and when AI got good my first feeling wasn’t excitement, it was that I was taking a math test with a calculator. If the machine can do everything I do, what am I actually for. Building the tools instead of just using one is how I answered that, at least for myself. When I build the system, the judgment about what to build, what matters, what’s wrong, that’s still mine. The craft stays mine. I’m not handing the thinking over, I’m pointing it at the parts that were never the point.

The pieces that turn AI into an amplifier

If I had to name the architecture in plain words, it’s three things. Memory, structure, and loops.

Memory is long-term context about my projects and my life, who I am, what I’m working toward, the decisions I’ve already made and why. It gets read at the start of every session, so the system doesn’t reset to a stranger every morning. This is the single highest-leverage piece, and it’s the one the chat window can’t give you, because memory you don’t control isn’t really yours.

Structure is my goals and my projects written down in a form the AI reads first, every time. Not a vague “help me be productive,” but the actual list of what I’m building and what done looks like. Once that’s in place, the suggestions stop being generic and start serving the thing I’m actually trying to do. Generic advice is the smell of an AI that doesn’t know you. Structure is the cure.

Loops are the part most people skip, and they’re my favorite. These are small jobs that run on their own, a morning brief, an activity tracker, little things that log to my phone, so the system is doing work while I’m not sitting in front of it. The first time something useful showed up without me asking for it, that’s when it stopped feeling like a tool and started feeling like infrastructure. I’m still figuring out the right shape for all of it, I won’t pretend I’ve got it solved, but memory, structure, and loops are the bones.

What it actually took to make it part of my routine

Here’s the honest part. Building it was not the hard part. Building is the fun part, the dopamine part, the part I could do all night. The hard part was using it every day, especially on the days I didn’t feel like it.

A system like this only compounds if you actually show up to it, and showing up is boring in exactly the way that matters. I read the first chapter of Atomic Habits right around the time I was deep in this, and it landed harder than I expected, because the whole project was secretly a habit problem wearing a technology costume. I had to make the using-it automatic, journaling daily, opening the system daily, feeding it the raw material of my days even when I had nothing interesting to say. Some days the only honest entry was that I had nothing, and I wrote it anyway, because the streak was the point and the streak is what makes the memory worth anything.

The failure modes nobody warns you about

The biggest one I hit has a name now, in my head at least. Fake progress. There were whole days where I’d spend hours improving the system, polishing it, adding to it, and get to the end of the night genuinely unable to tell if anything had actually moved forward. The tool had quietly become the procrastination. I was building the amplifier instead of using it to amplify anything, and from the inside it felt productive the entire time. If you build one of these, watch for this, it is the easiest trap in the whole thing to fall into and the hardest to see while you’re in it.

The second one is slower and scarier. If you let it do the thinking, not just the typing, the craft underneath you starts to hollow out and you don’t notice until you reach for the skill and it’s softer than it was. That’s the calculator doing the test again. The defense is boring but real, you have to keep doing the hard part yourself even when the machine offers to take it.

And the third is the dumbest and most common, trusting the output because it sounds confident. It is confident about everything, including the things it’s wrong about. I check its work. Every time it matters, I check its work.

How to amplify the craft instead of replacing it

The line I keep drawing is this. Point the AI at the mechanical parts, keep the parts that are judgment, taste, and voice. Let it transcribe, organize, fetch, scaffold, summarize, carry the boring weight. Keep the deciding. Keep the what-actually-matters-here. Keep the part that only sounds like you because you’re the one who lived it.

That’s not a compromise, it’s the whole strategy. The mechanical work was never the thing people valued, it was just the cost of admission to the thing they valued. If AI eats the cost of admission and leaves me the judgment, that’s not a threat, that’s the best trade I’ve ever been offered, as long as I’m disciplined enough to actually keep the judgment instead of slowly handing it over for the convenience.

Where I’m taking it next

The next thing I want is to deploy this at scale, to make it more autonomous without taking it off the leash, and to keep it honest as it grows, which is its own engineering problem. But I’ll give you the most concrete example I have of the whole thesis, because it’s this post.

This started as a month of handwriting in a paper notebook. I scanned it in, the system read every page, pulled out the threads, organized a month of scattered thoughts, and handed me back the raw material. But the words you’re reading, the choosing, the opinions, the voice, those are still mine. The amplifier didn’t write this. It just cleared away everything that wasn’t the writing, so the version of me that showed up here is the one that actually did the work. That’s the whole point. That’s what building your own AI is supposed to feel like.

If you want to build your own

I’m going to try teaching this. One weekend, live, a small group. You bring one part of your job that eats your time, and you leave Sunday with a working AI that actually runs it. Not a course you’ll quit at twelve percent done. The thing, running, by the end.

And I’ll be honest, because that’s the whole point of The Bellows: this is the first one, and I don’t fully know yet what it’s worth to someone who isn’t me. So I’m not asking you to buy anything today. I’m asking you to get on the list if you want this. If enough of you do, I’ll run the first cohort at a founding price, and we’ll build them together, in the open, like everything else here.