This will be my first post here, as recently I have rebuilt my website from scratch. Went on like a breeze, especially as I did it as part of my learning process of learning how Claude Code works and how totally mesmerized I got by Openclaw assistant I set up for myself!
I ain't gonna elaborate too much - if you're in tech industry, already you should know about Claude, Openclaw and amazing progress of AI evolution recently, along with agentic AI applications. Unless you've been sitting under a rock (or tied up to a legacy project) like me for last couple of years.
Setting up Openclaw assistant is refreshingly easy, if you have actual place, machine, where you can do it. I set up mine on inhouse small server that runs in the attic. For isolation, I set it up on it's own virtual instance, so it would not break anything else brittle, that I already have on the same physical machine. Virsh has proven more than convenient for such setup - I already have at least 3 virtual machines set up like this.
Once you set up openclaw and add it's gateway capabilities, like whatsapp, you can immediately start talking to it, almost to like a living person. Main, most powerful advantage it has over in browser prompt, either on claude.ai or chatgpt.com (despite some disadvantages) is that it can act independently, it can be set up to have access to any resources you can imagine. It can:
- run scheduled tasks
- deliver reminders
- produce reports
- access git repository of your choice
- access remote servers via ssh and run commands
- anything you can imagine, as long as you can give it tools for...
However it is easy to fall into a trap, as as convenient as it is, it may spiral out of control, if you're not careful. Here's my real life example...
This website, where you're reading this text, resides in a repository, to which my Claw agent has access to. It is hosted on VPS server, my Claw agent also has access to. One of the tools we (gonna say we, as me with claude code and/or with Claw, my assistant) implemented, was simple page visit tracking, bot detection and flagging unwanted visitors to block them.
At some point it turned out that detection job has a flaw, which caused me to be locked out from visiting my page from my local IP 😃 So, naturally, I asked my assistant what might have happened.
And all hell broke loose...
My assistant went on, proactive, trying to both debug and fix problem on it's own. While not having full context, and perhaps not having properly set up session when debugging, it made some false assumptions about state of the Puma app, running ruby version, etc.
Asking assistant what it did and what we can to do revert the state of things only started to make things... worse? As again, it went on guns blazing, making changes...
So that was lesson learned for me, to try and contain the assistant intentions by making instructions more clear. But hey! I am new to this stuff! 🙃
I thought it would be the end of it...
Nope.
Days gone by, but eventually some other issue happened...
Assistant once again imagined he's Clint Eastwood in Dirty Harry, went on guns blazing and shooting fixes all over the place... So I had to put him back to his place, make him write in memory what he's to do and what NOT to do.
After that, each time in similar scenario it started making suggestions and asking what to do next.
I know I may seem like an idiot here! But hey, I am learning and sharing my experience 🙃
Lesson learned: AI assistant is great tool, but as any great and robust tool, it can be dangerous. Like a swiss army knife: it has many applications and save your a lot of trouble, but you can still cut yourself... 😀