How I rebuilt the way I run my business in 2026.

The short version

Nine months ago I was a WordPress guy who had just started using Claude Code.

Today both of my companies run out of one folder of plain files that I own. It keeps my books, my pipeline, my content, my analytics and my client work. It watches what I do every fifteen minutes, logs what already happened on its own, and messages me when something needs a decision.

I call it my AIOS, and I didn’t set out to build it. It came out of one problem that kept getting worse, and then it kept going.

This page is the whole thing in order, with dates, because the order is the interesting part.

Before: I was the integration layer

On December 9, 2025 I built my first agentic workflow with Claude Code. Eleven days later I decided I was done building on WordPress, and two days after that I sat down and rethought what I actually do for a living, because it was pretty obvious the old version of my job was going away.

By January I was using Claude in the browser, a handful of Claude Code sessions, the mobile app, Gemini and ChatGPT. On January 24 I started building a “Personal OS” in Notion to hold it all together.

Here’s what that looked like day to day. Every one of those tools knew a slice of my business and none of them knew all of it. My deals lived in Close. My notes lived in Notion. My books lived in QuickBooks. Testimonials lived in Senja, my short links lived in a WordPress plugin, and my analytics lived in Google Analytics and then Plausible.

I was the integration layer. Every new chat started with me explaining my own business again, from memory, slightly differently each time.

On April 29 I published a video called You Need An AI Operating System. I had the idea before I had the thing.

The detour: losing everything I’d told them

In May I tried to solve it with a tool. I spent a few weeks with an agent tool called Hermes, and livestreamed it on May 1, 5 and 8. The ambition was right and the implementation didn’t hold up for how I work, so I went back to Claude Code.

That detour taught me the only thing that actually matters in this whole story.

I wasn’t losing the tools. I was losing everything I had told them. The context was the asset. The tool on top of it is replaceable, and it will get replaced, probably several times.

So the context had to live somewhere I own, in a format any tool can read, forever.

One folder

The first commit to my AIOS repository landed on June 13, 2026. My son was about a month old, so honestly a lot of this got built at strange hours.

It absorbed the half-finished Desktop folder I had started in May, the Hermes setup, and the Notion Personal OS. Everything moved into one structure:

• Entities - me, and each of my companies, with their own context and their own credentials.
• Connectors - thin code that talks to one outside service each: Stripe, Google, YouTube, my bank.
• Modules - small tools that do one job anywhere, like transcribing audio or rendering clips.
• Pipelines - the end-to-end jobs that know my world, like producing a podcast episode.
• State - one database for deals, clients, projects and money.

It’s all markdown and code. Claude Code reads it today. If something better shows up tomorrow, it reads the same folder and nothing is lost.

Two weeks, five tools gone

This is the part that surprised me.

• June 14: my CRM moved out of Close and into my own database. My link shortener moved off WordPress onto a tiny Cloudflare Worker.
• June 16: Notion was retired. I went through everything in it, kept the 51 pages that were actually worth keeping, and deleted the rest.
• June 19: testimonials moved out of Senja into my own store.
• June 22 to 24: QuickBooks. My bank feeds, Stripe and PayPal now get categorized into my own books automatically, which took my bookkeeping software from about $120 a month to about $1.50. I made a video about that one because nobody believed me.

Mission Control, the dashboard that sits on top of all of it, went live the same day as the first commit and got redesigned on June 27. It has 23 views now, from the sales pipeline to cash flow.

None of those tools were bad. They just each held a piece of my business hostage in a format I couldn’t reach, and once the context lived in one place there was no reason to keep paying for the pieces.

Rules, not prompts

July is when I figured out that the most important files in the AIOS aren’t code.

They’re the rules, and every one of them has a date and the mistake that caused it.

• June 22: a session confidently told a peer of mine that three real, shipped features were “speculative,” because it trusted old training data. Rule: anything that changes over time gets checked against a live source before it’s stated as fact.
• July 7: a session claimed it couldn’t read my task list, because a plugin it expected wasn’t installed. The code to read it was sitting right there. I removed every one of those plugins and made the plain connector code the only way in.
• July 10: I stopped saying “AI never writes anything without my approval,” because it was too blunt to be useful. It became three tiers. Record what already happened, on its own. Prepare what’s clearly coming, like a drafted email, and tell me it’s ready. Decide or send only with me. The short version: log and prep on your own, never send or decide alone.
• July 11: a session described one of my systems to me wrong, because it read an out-of-date README instead of the code. Rule: when the code changes, the docs change in the same task.
• August 27: a code comment claimed something that wasn’t true, and it got repeated to me as fact for a whole afternoon. Rule: no comments that describe what code does, ever. Only why.

This is the part I’d tell anyone starting their own. The rules are the system. A model is only as good as what it’s holding you to, and mine holds me to things I decided while thinking clearly.

It started watching

On July 11 I built the Watcher. Every fifteen minutes it looks at my work sessions, my email, my project boards, my folders, my commits, my calendar and my tasks, and writes down what happened.

Everything else reads from it.

• The board keeps itself current. When something objectively happened, like a client email, it logs it on its own. When something needs judgment, it messages me in Google Chat, and I can just reply to act on it. It has made 2,934 automated updates to my board since July 13, and 2,923 of them were plain record-keeping.
• My task list gets corrected twice a day, at 6am for today and 8pm for tomorrow. It replaced a morning email I retired on August 18, and the reason is worth stealing: the email was a report, it described things I already knew. The new one is a receipt. It changes the list and shows me what it changed.
• My analytics became mine. 536,059 rows across 15 websites, going back to October 2021, pulled out of four different analytics platforms, three of which I don’t use anymore. The platform is just a sensor now. The history lives in my database.
• It backs itself up every night, and checks that my client databases are backing up too.

There are 18 automations running as I write this, and every one of them shows up on a board with when it last ran and whether it worked. That board is the only reason I trust any of it.

It breaks, too. On August 19 one expired login took down every automation that uses the model at once, and sent me about ten separate alert emails, none of which said the one useful thing. Now it sends one message that tells me exactly what to do.

Pointed at the business

By August the system was good enough that I started pointing it at my actual decisions instead of my admin, and that’s where it paid for itself.

• I stopped my weekly livestream on August 27, because the numbers were clear. Streams earned 1.7 subscribers per hour on camera, and my regular videos earned 37.3. Seven streams produced one click to my site. Seven videos produced 28.
• I cut my offers from 25 to 6 on August 26, and my task lists from 8 to 3.
• 571 of 1,302 clicks on my short links were bots. I only know because I own the data.
• Every money number got one definition on August 29, after two tabs on my own dashboard showed me two different revenue figures for the same year. Now there’s one way to calculate revenue, one tax rate, used everywhere.
• I lost a build by giving away the diagnosis before the work was sold. That turned into a written playbook the same week.

The same way of working ships client work, which is the point of all this. The clearest example is the portal I built for Sheffler & Company, their entire proposal process as software, and version one shipped in 44 days.

What it changed in me

On September 7 the AIOS caught me being wrong about my own business.

For a while I believed content was where I hide, because the deal data said every 2026 win came from a relationship and none came from content. Then I noticed what that data couldn’t see. One year of deals can’t show relationships that took seven years of videos to form. Both things were true, and I had only written down the first one.

So the top-level constraint changed. The limiting factor is that not enough people know who I am.

That’s also the real reason this page exists. It’s the first time I’ve written the whole thing down in one place, and I’ll keep adding to it.

What it is right now

As of September 29, 2026:

• 1,743 commits since June 13
• 28 connectors to outside services
• 13 modules and 25 pipelines
• 18 automations running on a schedule
• 23 views in Mission Control
• 536,059 rows of analytics history I own

And it’s still just folders and files. Nothing about it is proprietary or locked up.

If you want one

You can build your own. That’s genuinely the goal, and it’s what I teach inside Modern Builders Society.

Or, if you run a business and you’d rather have the business version built for you, the one that holds your clients, your pipeline, your numbers and your operations in one place you own, that’s exactly what I do.