Everyone’s still clicking around in ChatGPT like it’s 2024. I get it, I was doing the same thing for a long time, bouncing between Claude and Gemini and GPT in the browser, copy-pasting stuff back and forth, losing context every time I switched. But somewhere around the end of last year something clicked for me and I realized wait a minute, I don’t need to be using AI in a browser at all.
My terminal is my operating system now. And honestly I don’t think I’m going back.
How It Started
It started with vibe coding. I was using Claude Code, which is basically Claude running in your command line instead of a browser, and I was building some projects with it. I built Streamliner.gg which helps live streamers generate titles and descriptions and thumbnails, and I built CustodyJournal.com which helps fathers keep track of everything during custody battles because trust me I know how messy that gets.
And as I kept using Claude Code for these projects I started noticing something, this thing can do way more than just write code for me. Every week they’re adding features like MCP servers and skills and custom agents and I’m sitting there going okay what if I stop treating this like a coding tool and start treating it like an actual employee that knows everything about me and my business?
The Folder That Runs My Life
So I set up this folder on my hard drive called AiMe. It’s backed up to GitHub and to my Synology NAS so I’ve got redundancy, but basically this one folder holds the entire context of my life in a bunch of markdown files.
Here’s the structure. I’ve got an inbox for quick ideas and thoughts, a brain folder which is the big one, an LLM context folder that tells AI how to work with me like what I like and don’t like and how I communicate and what my goals are, a projects folder, templates, outputs, and an archive because I never delete anything.
The brain folder is where it gets wild. It’s got my journal entries, health information, legal stuff around my custody battle with my son Tanner and my daughter Anna Mae, personal development notes, and then there’s the Omi transcriptions. I wear this device called Omi around my neck and it records and transcribes everything, every conversation, every meeting, everything. There are almost 2,000 transcription files in that folder alone. All of that gets indexed into a RAG server, which is basically a searchable memory database for the AI, so when Claude Code needs to find something about me it can search through years of context instantly.
The Daily Briefing That Changed Everything
Every morning I run a custom command called /today and this thing is honestly one of the best parts of the whole system. It generates a daily briefing by looking at my projects, my task list, my calendar, everything. It asks me if there’s anything specific I want to focus on for the day and then it puts together a prioritized plan.
But here’s the part that really sold me on it. Before it finalizes anything it sends the plan to what I call the Employee Council, which is basically five AI personas that review the plan and score each task. They’ll say stuff like “recording the AI video scores 8.1 out of 10 for growth” or “switching your internet provider is only a 5.5” and they’ll call out blind spots like hey you said you wanted to follow up with that sponsor three days ago and you still haven’t done it.
It keeps me accountable in a way that a regular to-do list never could because it has context on EVERYTHING I’m supposed to be doing, not just what I remembered to write down.
The Autonomous Employee
The craziest part of this whole setup is what I call the employee skill. When I activate it Claude Code goes into autonomous mode where it just scans my projects and context and identifies work it can do for me without me even asking.
It wrote blog posts for two of my YouTube videos by pulling the outlines and scripts and context from my projects folder and combining that with my writing style guidelines. And these aren’t lazy first drafts either because the employee sends its work to that brutal council I mentioned for feedback and then iterates on it until both the employee and the council agree it’s good enough.
It creates social media content too. Twitter threads, LinkedIn posts, YouTube community posts, Reddit cross-posts. All formatted for each platform in particular. And it just drops all of this into my review queue so the next time I run my daily briefing it says hey you’ve got five pieces of content waiting for approval.
While the employee is running I can open another terminal tab and work on something completely different. It’s literally working in the background like an actual employee would.
16 Custom Agents and Counting
Over time I’ve built up 16 custom agents that each handle different parts of my life and business. There’s a Brutal Critic that destroys my YouTube scripts using a specific framework before I record them. A Chat Processor that takes conversation files and Omi transcriptions and organizes them into the right folders with proper metadata. A Content Repurposer that looks at existing content and figures out what else it can turn into. A Context Miner that does weekly strategic reflections and surfaces patterns I might not have noticed.
There’s a YouTube Comments agent that surfaces comments worth responding to and extracts content ideas from what people are saying. A Weekly Analytics agent that pulls data from YouTube and all my social media accounts and gives me insights. A Research agent. A Session Closer that saves everything and commits to GitHub when I’m done working for the day.
The thing is I didn’t plan all of these from the start. I’d be using the system and notice something like man it would be nice if I didn’t have to manually go through YouTube comments every week, and then I’d just tell Claude Code hey let’s build an agent for that. And when an agent isn’t working right I just tell Claude Code what I didn’t like and it updates the agent. The system keeps getting better because I keep using it and feeding it feedback.
Why Owning Your Context Changes Everything
The biggest thing for me is that I own all of this context. It’s not locked inside ChatGPT’s servers or Claude’s cloud or Google’s infrastructure. It’s sitting on my hard drive in text files that I can back up and version control and move wherever I want.
That means if I want to use Claude for a task that needs strong tool use I can do that. If I want to use Gemini for creative writing I can switch and Gemini has all the same context through the RAG server. If I want to use Codex because it’s better at a specific programming language I just open it up and it already knows everything about me and my projects. Brand new chat, zero prior conversation, but full context from day one because the context lives in MY system not theirs.
I went from using Windows as my operating system and AI as a tool, to using AI as my operating system and Windows as just the thing that runs underneath it. Claude Code isn’t writing code for me anymore. It’s managing my entire life, my business, my content, my schedule, my relationships, everything. And honestly the really cool thing is we’re still in the early days of this stuff, so the system I have now that already does all of this is only going to get more capable as these models improve.
The Real Cost Breakdown
People always ask how much all of this costs and I think they expect some crazy number but it’s honestly not that bad. Let me break it down because I was genuinely surprised when I ran the numbers.
Claude Code is $20 a month for the Pro plan which gets you Claude 4.6 Sonnet. If you want Opus which is their strongest model that’s $200 a month with the Max plan but honestly Sonnet handles 90% of what I need it to do. I use the Pro plan for most of my day-to-day work and only switch to Opus when I need really deep reasoning on a complex problem.
The RAG server is free because it runs locally on my machine. PocketBase which I use for task management is free and open source. GitHub for backups is free for private repos. The Omi device was a one-time purchase of $89 and the transcriptions come with it.
So my entire AI operating system costs me $20 a month at the base level. Compare that to hiring a virtual assistant at $500 to $2000 a month, or a content writer at $50 to $200 per blog post, or a social media manager at $800 to $3000 a month. I’m getting the equivalent of multiple employees for the cost of a Netflix subscription. The ROI is absurd when you think about it.
The expensive part is the time investment in setting it up. I’ve spent probably 200 hours over the last few months building out the system, writing agent prompts, configuring MCP servers, organizing my context files. But that time investment compounds. Every hour I put into the system saves me multiple hours going forward because the agents just keep running.
What Went Wrong Along the Way
I don’t want to make this sound like everything was smooth because it absolutely was not. There are some real pitfalls that I want to be honest about.
The biggest one was context pollution. Early on I was dumping everything into my brain folder without any organization and the AI would pull up random journal entries when I asked it about business strategy, or it would reference a casual conversation I had with a friend when trying to write professional content. I had to restructure the entire folder system and create clear boundaries between personal context and business context. That took a full weekend and I should have done it from the start.
Another issue is hallucination. When the AI has access to a lot of context about you, it can sometimes confidently reference things that never happened or combine memories from different events into one. I’ve caught it telling me I said things in meetings that I never actually said. The Omi transcriptions help because they’re ground truth, but you have to stay vigilant. Always verify before you act on something the AI tells you about your own past.
Rate limits are real too. When I’m running multiple agents simultaneously they can hit API limits, especially on the $20 plan. I’ve had agents fail halfway through a task because they ran out of tokens. The workaround is staggering your agent runs and being strategic about which tasks actually need the most powerful model versus which can run on a faster cheaper one.
And honestly the biggest risk is over-delegation. There were weeks where I let the AI handle so much that I lost track of what was actually happening in my own business. I’d approve content without really reading it, let agents make decisions I should have made myself, and just kind of checked out. That’s a trap. The AI is an employee, not a replacement for you. You still have to be the CEO.
Where This Is All Going
I genuinely believe that within two years the way most knowledge workers interact with computers will look more like what I’m doing now than the browser-tab-and-mouse approach that’s been standard since the 90s. The CLI-first AI workflow isn’t a geeky experiment anymore, it’s a productivity paradigm that’s going to go mainstream.
Computer use agents are getting better every month. MCP servers are becoming standardized. More tools are adding CLI interfaces in particular designed for AI integration. The ecosystem is growing fast and every new tool that plugs in makes the whole system more capable.
What excites me most is the compounding effect. My system today is exponentially more capable than it was three months ago, not because the AI got smarter but because I’ve built more context, refined more agents, and connected more tools. Three months from now it’ll be even better. A year from now I legitimately can’t predict what it’ll be doing for me because every time I think I’ve hit the ceiling something new opens up.
If you’re thinking about trying this yourself, start small. Pick one CLI tool. Set up a folder with some basic context about yourself and your work. Build one agent for one repetitive task. See what happens. You don’t need my setup to get value from this approach, you just need to stop thinking of AI as a chat window and start thinking of it as an employee who works in your terminal.
If you’re still using AI exclusively in the browser you’re leaving a lot on the table. Start with a simple folder structure, pick one CLI tool whether that’s Claude Code or Codex or Gemini CLI, and just start building context. The system doesn’t have to be perfect on day one. Mine wasn’t. It just has to grow with you.