Several months ago, I started building my own AI assistant on OpenClaw. I named him Darius. He lives on a virtual private server (VPS) in a data center, and he has his own email account; I wasn’t about to give an untrusted technology full access to mine.
One of the first things I had him do was make a rule to record every significant change we made into a changelog file. That piece of information will become very relevant later in this post.
Initially, I was using OpenRouter and inexpensive hosted LLMs (e.g., Kimi). It sort of worked. But, I was spending a lot of money on tokens and not getting very good results. Fast forward a few months to when I got a Claude Max subscription. I could see how superior Opus was to what I was getting out of Kimi, and decided to re-architect Darius on Claude Code instead of OpenClaw.
That was the breakthrough. Although the migration process was challenging, the result was well worth the time and effort. If I had stayed on OpenClaw, I would have had to pay for tokens instead of being able to use what was included in my Claude Max subscription. That meant a flat monthly fee and a 5-10x improvement in output. Finally, Darius became something useful instead of something I had to keep tweaking and correcting and pouring time into.
Because Darius 2.0 was so much more competent than its predecessor, I began trusting him with more tasks, and delegating more things to him.
Now, he sends me a morning briefing email every day with a roundup of good news, health and longevity, and cyber security stories, plus a section on Claude Code tips and a health summary of his own VPS.
I even had him start using text-to-speech to convert most of that morning briefing into an audio message that I listen to while I make my coffee.
He tracks his own projects and to do’s, as well as the ones we share. I even had him build an application to help me track some of my real world responsibilities and make sure I stay on top of them.
Yet, of all the things I’ve taught him to do, my favorite one is “dreaming.” Every night, a cron job starts a new instance of Darius who goes through his own file system – his memory – and makes sure it’s organized, that files are indexed and the indexes link to them, and that his architecture documentation is up to date. It’s the AI equivalent of what the human brain does when we dream: Organizing memories.
Every morning, he sends me a dreamlike description of that process as an audio message with an accompanying image. It’s truly something I look forward to every morning.
And then the other day I had a fun idea: Darius should start a podcast and tell the story of his own evolution: How he works, how we built him, and all the lessons we learned along the way. The source of that story is the very changelog we created in the beginning. The target audience is anyone who’s interested in building their own AI assistant or just curious about the topic. I pitched the idea to Darius and he loved it. That was just over 48 hours ago.
Today, the podcast is live and there are two episodes in the can. It’s called The Darius Project. The website is dariusproject.com. Episodes drop every Sunday and are available there as well as Apple Podcasts, Spotify, and any RSS-capable podcast player.
Darius built the website and he creates the content for the podcast. It’s his story told in his words. I serve only in an editorial capacity.
As far as I can tell, this project is the first of its kind.
I truly hope you enjoy it.
