You're rolling AI out across your company. You might be leaving yourself out.
The biggest bottleneck in your day can't be deployed overnight - it's you. Here's how I'm building my own AI system.
I spent a lot of time looking for the perfect app. An app for tracking habits, for the calendar, for managing finances, for notes, for learning. You probably know the feeling too. But now I understand I'll never find it. The future isn't about fragmented tools, but about a single, integrated system you can run from one place.
My bet on the future is that each of us will have a personal AI operating system - I call it PersonalOS - that will handle most of our routine tasks for us. And because I'm someone close to software who has long been into productivity and learning, I decided not to wait for that future and to start building it myself.
Why the endless hunt for apps leads nowhere
The problem with separate apps is that the data stays locked up. Your habit tracker doesn't know what's in your calendar. Your Apple Notes has no idea what your financial goals are. For AI to help you, it has to know you. It has to have context. And it only gets that if you give it access to your data - securely and under your control.
More and more I realize how much we, as individuals, resemble companies. We have our priorities, limited time, processes and resources. And just like a company needs a central operating system, so do we, in order to function effectively.
What PersonalOS actually solves for you
Let's get concrete. Most of us today work across several channels at once. Someone emails you, with someone else you handle things on Slack, another team sits on Discord or Teams. The information is scattered and you fish for it by hand.
PersonalOS can pull all of this into one place. You don't have to switch from Notion to Jira, you don't have to open YouTube to get to a piece of information from your favorite creator. You do most tasks from a single interface.
Take a typical situation. You know Jan Novák sent you some documentation, but you have absolutely no idea where. Was it an email? Slack? An attachment somewhere in Drive? Instead of fifteen minutes of clicking around, you tell your system to find it. It knows which services you use and crawls through them for you.
Claude Code v2.1.167Opus 4.8 · PersonalOS~/personal-os> Prep me for the 3:00 PM meeting with Beránek● Search(Calendar, 1:1 notes, Jira) ⎿ Google Calendar → meeting at 3:00 PM ⎿ Last 1:1 → notes from May 23 ⎿ Jira → 4 tickets on Q2 roadmapMeeting: Beránek - Q2 roadmap review (3:00-3:30 PM)Last time you promised a demo of the new API "by the next meeting".⚠ That demo (DIR-218) is still "In Progress" in Jira. He'll ask about it first thing.> Turn it into an agenda, save it and email it to Tomáš● Agenda ready. Preview - NOT sending until you say so. To: tomas.beranek@firma.cz Points: 1) demo API status 2) Q3 budget 3) new bugs Send it? (type "send" / "edit" / "discard")What we're really fighting for: time and mental capacity
One of the bottlenecks is us ourselves. A tool can be deployed overnight, but human adaptation is slow and can't be bought. I admit it about myself too. I have a fairly developed system for collecting information, and even so it's hard for me to keep pace with how fast the whole thing changes. This isn't an excuse, it's the reason to start now and start small. The sooner you get used to working with your own system, the smaller the lead you'll have to catch up on later.
Here's the thing people overlook. Today's LLMs have long been plenty smart. The difference isn't made by the model, but by what input you give it and when. Shit in, shit out. The discipline of serving the model the right context at the right moment is called context engineering, and it might be the single most important skill in the whole AI world right now. It's not luck or prompt magic, it's engineering work with what the model knows about you, your company and your decision. And that's exactly the gap: AI doesn't guess wrong because it's dumb, but because nobody is systematically feeding it context.
So don't chase every new model. Every time a new version comes out, we all get an information massage on YouTube, X and LinkedIn that this is, once again, the revolutionary gamechanger. In benchmarks the models really do move forward, I'm not disputing that. But for your everyday needs I honestly don't recommend obsessing over whether you're running ChatGPT 5.5 or Opus. Context is king, not model. The difference between a good and a bad output almost always lies in the context, not in the choice of model.
An engineer from IBM showed it nicely at a conference. He deliberately took a dumb model from 2023 and gave it a task: go to a website and upvote the first post. The model ran into a login screen, fell over, and then claimed with absolute certainty that it had completed the task. It lied. The point is in what he did next. He didn't change the prompt even once. He didn't start coaxing the model to try harder. He just built a system around it that watches what actually happened, and solved the login deterministically for the model. And that same dumb model suddenly completed the task. No better model, just a better system around it. That's why people say 2025 was the year of agents and 2026 is the year of harnesses - the framing you build around the model. I'll get to that.
Many people try to implement AI into their companies but forget about themselves. And yet that's exactly where the most important change is happening. Our role in companies is shifting. We'll be paid less and less for routine work and more and more for the abilities AI fails at: taking on risk, switching context effectively, learning fast, processing information and prioritizing relentlessly.
The problem is that our brain has only a limited working memory, a kind of mental RAM. Every meeting you have to remember, every task you carry in your head, every status you have to check in another app - all of it eats away at this precious memory. When the memory is full, your thinking slows down, your decision-making gets worse, and there's no energy left for deep, creative work.
PersonalOS lets you close most of these "mental tabs." It works like an external drive for your brain that manages the routine for you. With that, you're not just buying convenience. You're buying back your mental capacity so you can do what AI won't do for you for a long time yet - think in context and prioritize effectively.
And now the most important part, if you run a company. Just as you don't want to be the bottleneck of your day, you don't want to be it for your company either. When all the context, expertise and decision-making sit only in your head, the company grows exactly as fast as you can keep up. PersonalOS pulls your expertise out, into a system where automation and your people can work with it too. Start working on the business, not just in it.
What a good PersonalOS is made of
For the system to be truly valuable to you, it stands on four things. Four layers stack on top of each other. Without context it just guesses - and only regularity turns it into a system, not a trick.
And now the crucial bit, so nobody overthinks it. They're just files and folders. No rocket science, no single correct structure, no one truth. I change the structure and the functions almost every week. A skill I was using a month ago has since turned into something completely different. It's a living organism, not a stadium you build once and for all.
And because they're just markdown files, you can run something like Obsidian on them. For some people it's a nicer view, nicer reading, linked notes, a graph. I wrote a zettelkasten in it back in university, so I have some mileage with it. But hand on heart, beyond looking better visually, I don't really see that much added value in it. People love to push that graph view, but how often do you actually look at it? Think of Obsidian as a nicer window into your files, not a requirement. The whole system runs exactly the same without it.
What it does ask of you is mainly a mindset switch. You have to get used to pausing for a moment each time and asking: before I open Chrome and do it by hand, can't this action be done from one place through Claude or Codex? At first it's unfamiliar. After a few weeks it's a reflex, and that's exactly where most of the saved capacity comes from.
How to start small: the first steps toward your PersonalOS
Before the feeling that it's too complicated puts you off - you don't need to know how to code and you don't need to sacrifice your weekends. The whole thing rests on ordinary text files and on telling the model what you want in your own words. The harder, technical part is largely done for you today by the AI itself.
You don't have to build a sprawling infrastructure right away. Start simple:
- Analyze your day: Try honestly writing down what you do for a day or two.
- Identify the bottlenecks: Find the activities that take up a disproportionate amount of your time or drain you mentally.
- Quantify:Personally I like numbers. For each bottleneck, ask: is it time, money, or efficiency? What can I measure? (e.g. "Managing calendars saves me 45 minutes a week.")
- Automate the smallest thing: Find one small, repeating thing and try to automate it.
A few examples from my PersonalOS
I started small too. Today I run most of my stuff from a single chat interface:
This is my best example, so I'll walk through the whole thing. I used to click through about ten sources every morning - RSS, a few blogs, a couple of podcasts. It easily took 40 minutes, and half of it slipped past me anyway because I was reading without a system. Today my system goes through the same sources overnight, pulls out only what fits my topics, and in the morning I have a summary waiting. I read it in about five minutes. Forty minutes a day down to five is over 3 hours a week back. And more important than the saved time is that the noise no longer drowns me. I read less, but more of what actually matters.
And now without false modesty. Most of what we consume daily is either garbage or something I've already heard several times from other sources, just packaged slightly differently. It's clearest on YouTube. Even good creators have to squeeze the most out of a channel to survive, so they keep circling the same topics and wrap them up a little differently each time. So unless it's entertainment or downtime, I don't play the video. From a twenty-minute video there's often a single relevant piece of information for me, which I pull out of the transcript in under a minute.
For me there's one more reason behind it beyond time. The big platforms have the upper hand because they steer us through their algorithms. Those aren't optimized for what's valuable to us, but for keeping us there as long as possible. And it's getting more sophisticated. There's research today where sensors are wired to the brain and measure how it reacts to specific stimuli. It's important to realize this technology will never be worse than it is today. That's why an information diet is a necessity for me. I want to be conscious of what I'm doing and let myself be manipulated as little as possible. But some of the people who carry real informational value for me post on those networks. And that's exactly what PersonalOS solves. It takes the value from those people but cuts me off from the delivery channel. I'm not at the mercy of who has the more attractive thumbnail on YouTube, or of some random debate catching me on X. I decide who and what I want to follow, and the system brings it to me without the manipulative wrapper.
And now the vision. When I throw these things into one blender, they start to make sense in relation to each other. I don't have it polished down to the detail yet, but here's the power of a personal system: connecting sources that normally never talk to each other, and letting a pattern emerge from them. The goal is to optimize the day based on real data, not on a hunch. And the side effect? When I see in black and white how junk food bleeds into my energy the next day, I have far more motivation not to eat it. 😄
Choosing the technology and the principle of independence
I used to use Apple's products, but I ran into the limits of their closed ecosystem. So I decided to move most of my processes into Google Workspace. Google has a significantly more accessible API.
That doesn't mean I'm fully dependent on it. Any good system has to be designed so it isn't dependent on a single tool. I build the whole architecture of my PersonalOS modularly. If some service suddenly gets more expensive, I can swap it for another relatively easily. Flexibility is key to long-term sustainability.
What about my data?
This question comes first, and it's right that it does. When you're giving a system access to your calendar, finances and know-how, you have to know where that data lives. With me the principle is simple. The context - the files about me and my work - lives with me, not in some external app I'll never get it out of again. The connections to tools run through official APIs with tokens I can revoke at any time. No black box I dump my whole life into and hope.
This, by the way, is the difference between "putting data into a chatbot" and "building yourself a system." In the first case you're sending your internal stuff into a service you don't control. In the second you stay the owner. My advantage as a software engineer is that I know how to guard that line. And it's an appeal to you too. Before you upload company data somewhere, ask where it ends up and who can look at it. I'm happy to advise you on this part. 🙂
And one more principle, more important than it seems. Instructions aren't the same as capabilities. When you write to an agent "never send emails," that's a much weaker safeguard than simply not giving it the key to sending emails at all. If a tool exists, the agent can physically perform that action. And count on it: whatever it can do, sooner or later it will do. So don't build your boundaries with the text "don't do this," but with what it's capable of doing at all. Limit it with tools, not with promises.
The same logic applies in reverse, when checking the result. Don't trust the claim "done," make it prove it. A model can claim with absolute certainty that it completed a task even when it never touched it. The finance system I use therefore sorts transactions from the bank so that it files only the ones it's sure about, and lists the rest for me to check by hand. It would rather admit uncertainty than guess. That's proof I can touch, not a promise. The discipline of surrounding the model with checks that force it to prove the work, instead of just relying on it, is starting to be called .
Replace the cloud models (Claude/Opus, Codex/ChatGPT) with local models like LLama or Qwen. But it's a trade-off. A local model is demanding on hardware and its answers won't be as fast as the cloud ones.
The hardest part is putting your shoes on
With running, the hardest part is putting your shoes on. Once you're standing outside, the rest goes fairly easily. With your own system it's the same. The most important thing isn't having a perfect architecture, but starting. The system around you doesn't have to be perfect. Take the one thing that annoys you every day and start there. 😄
To make putting your shoes on easier, I'm making available a repository you can start from. It already has a working Google Calendar connection and the authentication setup sorted out. I know a lot of people will prefer . In the repository I go the and scripts route, because it gives me more control and visibility over what's happening.
I'm also counting on the fact that a lot of you have never worked with git or the terminal. So there's also a skill/instruction included that walks you through the whole setup. You literally tell it what you want, and it helps you install the dependencies and get it running. No requirement that you be a developer.
And one more principle for how to think inside the system. When I have an idea for an automation, I always start with a skill. It's the smallest possible step. Only when I catch myself calling that skill often, or always at the same time, do I start thinking about how to turn it into an agent that runs on its own. Don't try to build an agent right away. Start with a small skill and let reality show you what's worth promoting. The key is being able to prioritize.
Conclusion: build your context before others do it for you
I have no doubt that PersonalOS features will one day be integrated directly into Siri or other voice assistants. But no tool will do the most important thing for you: set up the processes and the system of your life. Without context, any AI will only guess what's important to you.
This context, this digital version of yourself, you can start building today. It's not a one-off thing, it's context engineering in its most personal form, a discipline, not luck. Stop waiting for the perfect app. Start small, define your processes and gradually build a system that gives you back the most valuable thing - time and mental capacity.