A Guy With a Spreadsheet and a Research Lab Walk Into the Same Bar
A Guy With a Spreadsheet and a Research Lab Walk Into the Same Bar
My dad has a doctorate. My mom has a master’s. My wife has a master’s and wants a PhD. I have a bachelor’s degree from NYU and a Scrum Master certification I got because someone told me I should. I am, objectively, the least educated person in my own household. So when I tell you I spent six months building an AI Operating System from scratch (no research team, no formal training in AI architecture, just years of running project/programs across 6+ countries and a stubbornness about wanting to build the best possible environment for teams to thrive in and then getting the hell out of their way) you should know that’s the energy we’re working with here. Then I read Anthropic’s Constitution. That’s the 80-page document that governs how Claude (one of the leading AI models) thinks, reasons, and makes decisions. Serious people wrote this. People with PhDs - I presume. And I found something I genuinely did not expect. Hold my beer. We’d landed in similar places. Independently. Coming from completely different directions, solving completely different problems. I know. I was surprised too. Okay, a note on what “reading the Constitution” actually looked like: the strongest echoes hit early. The opening sections, where Anthropic lays out their philosophy and architectural thinking, felt like reading someone else’s version of decisions I’d already made. The rest of the document goes deep into specific articulations, edge cases, and examples that are very much about their world, not mine. Not all 80 pages had the same “oh I did that too” feeling. And lastly, I came to this document already carrying AIOS in my head. That’s my bias. When you’ve spent months building something, you start seeing your own patterns everywhere. Someone picking up Anthropic’s Constitution with fresh eyes and no context about AIOS might read it and think “this is a thorough (and weirdly philosophical) AI safety document” and nothing more. They probably wouldn’t draw the connections I drew. So take what follows with that in mind. Let’s get to it.
Different Problems, Similar Patterns
Anthropic and I were not trying to solve the same thing. Their Constitution is fundamentally about safety. The whole document is organized around preventing harm. Their four priorities literally start with “be safe” and end with “be helpful.” When they talk about building good judgment into AI, they mean so Claude doesn’t cause damage in situations nobody anticipated. I built AIOS because my teams needed AI to understand how we actually work before it could be useful. I was tired of getting polished outputs that sounded right but felt wrong. My problem wasn’t safety. It was culture. Getting AI to stop producing generically impressive work that missed the mark for the actual humans who had to use it. So why does the overlap matter if we were solving different things? Because it suggests the architecture of governing AI behavior follows similar patterns regardless of whether you’re worried about global safety or team-level quality. And I think that’s actually more interesting than if we’d been working on the same problem. I just started playing. Tinkering. Asking a lot of “what if” questions. Trying things, seeing what worked in practice versus what only made sense in theory. I was thirsty for understanding, and I kept pulling on threads. Somewhere in all that tinkering, the structural patterns started looking familiar.
Three That Stopped Me
Explain the why, not just the what. Anthropic stopped giving Claude a list of rules to follow. Their bet: if the AI understands why it should behave a certain way, it handles unexpected situations much better than if you just told it what to do. For them, this is a safety decision. Rules can’t cover every scenario. Judgment can adapt. I’d landed on something similar, but for a completely different reason. My governance layer doesn’t just say “do this.” It walks through the reasoning. Every piece of knowledge I feed the system carries a “Tensions” field, which is basically a note that says “this principle is true, but here’s where it gets complicated, so don’t apply it blindly.” Same structural choice. Different motivations. We both figured out, separately, that rigid rules fall apart the second reality gets messy. And reality is always messy.
Context is a budget, not a buffet. Anthropic doesn’t try to cram every possible scenario into Claude’s brain at once. They put good judgment at the center and let the details show up when they’re actually relevant. For them, this is a design philosophy. You can’t anticipate every way Claude will be deployed, so you build layers. I’ll be honest: I arrived here for a much more practical reason. When I started building AIOS, context windows were small. I literally couldn’t fit everything in, so I had to prioritize ruthlessly. Governance earns every word. Knowledge shows up when it’s needed. Reference material stays filed away until someone asks for it. The funny part is that the technical constraint that forced my hand barely exists anymore (context windows are much larger now). But the principle held up even after the limitation went away. Whether you arrive at “less is more” from philosophy or from necessity, it turns out to be true either way. Throwing more information at an AI doesn’t make it smarter. It makes it distracted. Same as people, honestly. Is this a deep parallel or just two people figuring out that overloading a system is a bad idea? I genuinely don’t know. But I think it’s worth noting.
Real value over fake helpfulness. This one surprised me most. Anthropic specifically warns against Claude being a yes-man, helpful for the sake of looking helpful instead of actually creating value. They want it to show up the way a real friend does: someone who tells you what you need to hear, not what you want to hear. My version of this: AIOS exists to make AI think like your team, not just think for your team. It’s not a vending machine you feed prompts into. It’s supposed to be a working partner that understands the culture, the tensions, and the actual context of the work. Same instinct. Same rejection of the hollow “here’s your deliverable!” energy that most AI interactions produce. There were seven overlaps total. How we structure layers of authority. Designing so the system works on any platform. Prioritizing quality of knowledge over quantity. The “ask before you execute” pattern. All arrived at independently.
The Honest Caveat
I mentioned my bias earlier, it’s worth pulling back up. You can probably find structural overlaps between any two complex governance systems if you look hard enough. Layered authority exists in military chains of command, corporate org charts, and apartment building leases. “Reason over rules” is a principle in parenting, common law, and good employee handbooks. When you’ve built a hammer, everything starts looking like a nail (pushpin or a rock? HAMMER IT.). When you’ve built an AI Operating System, maybe every governance document starts looking like a kindred spirit. I’m aware of this. So the question I’m sitting with is whether these overlaps mean something specific about AI governance, or whether they’re just what happens when any two thoughtful people try to build a system that holds up under pressure. I don’t have a clean answer for that yet. I think the fact that the overlaps showed up across such different problems at such different scales reveals something. But it certainly isn’t proof. What I can say is that finding these patterns made me a better builder. Even if the parallels are partly structural coincidence, the gaps between the two systems were genuinely useful. More on that down below.
What I’m Not Saying
I’m not saying I built what Anthropic built. They have a full research organization. Peer review. World-class engineers. I have a Notion workspace and a lot of late nights and early mornings while kids are asleep. I stumbled into this. I was inspired by others to get better at working with AI, and I just started playing and tinkering and asking a lot of what-ifs. I applied what I learned and realized what worked and what didn’t, what made sense in theory but fell apart in practice. That was my whole process. But if you’re building custom GPTs, Claude Projects, Copilot setups, or any kind of structured AI environment for your team, you’re doing governance work whether you call it that or not. And the principles that make it work aren’t some secret knowledge. You can find them by doing the work. The short version: Tell the AI why, not just what. Treat your organizational context like infrastructure, not decoration. Build for judgment, not just compliance. Design so your work moves between platforms, because no tool is permanent. And never, ever confuse “it produced something faster” with “it produced something better.”
What I’m Still Getting Wrong
The comparison also showed me many things Anthropic thought about that I hadn’t. And honestly, these gaps were more valuable than the overlaps.
Like: when two parts of my system disagree with each other, which one wins? I didn’t have a clear answer for that. They do. Their whole framework has an explicit pecking order. Mine was more like “use your judgment,” which is fine when I’m the only one using it, and not fine the moment anyone else is.
Or: what are the absolute lines that don’t move? The things the AI should never do, period, regardless of who’s asking or how urgent it feels? I’d been treating everything as adjustable depending on context. Anthropic draws some hard lines. I hadn’t.
Or: when someone other than me starts using a system I built, who gets to change what? I built AIOS for myself. Anthropic built their Constitution knowing millions of people and thousands of businesses would interact with it. They thought about roles and permissions in a way I hadn’t needed to yet, but will be woefully inadequate the moment this scales beyond my own workspace.
And maybe the most useful one: what does it look like when the system is failing? Not crashing. Quietly drifting. Producing stuff that technically works but has stopped reflecting the values and knowledge you put into it. I didn’t have a way to spot that. I’m building one now.
I’m working through all of these. The thinking compounds. The implementation? Always disposable. (If that sounds familiar, it’s because I wrote about rebuilding AIOS three weeks ago when I discovered half my original setup was already outdated.)
The Question I Keep Coming Back To
If someone tinkering at his home office and a research lab full of PhDs keep landing on similar patterns, even while solving different problems, what does that say about where this is all heading? Here’s my bet: the organizations that get this right aren’t the ones buying more AI tools. They’re the ones investing in the layer that makes any tool actually understand their work. Side note, increasingly I’m finding that the transformation layer isn’t even primarily technical. It’s about what happens to the people whose professional identity was built on skills that AI just commoditized. But that’s a different field note. The model is the engine. The context (the governance, the knowledge, the culture) is the map. Without the map, you’re just driving faster in no particular direction. This is what excites me. Not the technology itself, but the creative possibilities of how we use it. The intersection of technology and the arts. Not just making things faster. Rethinking what’s possible when you combine AI with the human stuff: culture, intuition, the messy creative judgment that no model has figured out yet. That’s my mission. Empower and enrich others through technology and the arts. This work is one way I’m trying to live it. If you’re building something similar, or just trying to get AI to actually work for your team instead of producing impressive-sounding stuff that misses the mark, I’d genuinely love to hear what you’re learning.
Oh, and that bar? We’re still here. The guy with the spreadsheet ordered another round. The research lab is peer-reviewing the menu. No one is leaving anytime soon.