AI isn't so bad...
Yes, AI is everywhere. And not always in a good way. Full disclosure: its not here. This blog is Ghost and nothing here is written, scripted, reviewed, or edited by AI. I also totally understand the ridiculous amount of things which have AI baked in now and have no business doing so. I especially don't like it forced upon me in my personal technology without my consent.
However, that does not mean I don't use AI.
When I say "use" I have to be specific. 99% of people think of ChatGPT on their phone when they think of AI. While that is super convenient and I use it a ton (having fully divested myself of Google searches now), its no longer the way I use AI most frequently.
Agentic AI
I have discovered agentic AI. Most people have never heard of this term despite being very informed on the subtle nuances between ChatGPT, Claude, and Gemini. Agentic AI is used for people who want to build things.
Many people don't need to build things, but I think you would be amazed once you start using it how many things in your life you may want to create. For example, do you do something repetitively every day? What about things which are hard to read because the UI is awful? Or how about being able to visualize things from different places all on one screen?
I created one thing. Then I started seeing many other places in my digital life I have been unhappy and have just been ignoring it. Now I build lots of things.
But What Is It?
All of the big AI providers have a command line version of their interface which can operate agentically. I have used them all, and I will tell you Claude Code by Anthropic (at the time of this writing) is head and shoulders above the rest. I use the Max plan ($100/mo) which means I have been using the Opus 4.5 model.
So when I say "agentic AI" I am referring to Claude Code (Opus 4.5) running in the command line of my computer. Its not a local LLM, which means I don't need a powerful computer since all the computation crunching is being done on Anthropic's servers. And it has no pretty interface which runs in your web browser. This is the old school console with the black background and white text and no picture - pure rock and roll.
How Does It Work?
What makes this different than something you run on your phone is Claude's ability to read, write, and self-correct. For example, I run Claude Code in Proxmox LXC containers so everything is fully sandboxed (not today Skynet). Within that container, I will start Claude in a /project directory with a context.md file which is basically a really long prompt with a ton of context created by AI (check out Prompt Cowboy).
Here is where the magic happens. Within my project, Claude can now write code directly - no copying and pasting. It can then execute that code. Then it can read the logs and see if there is an error. If there is? It can fix it itself. No human intervention necessary.
This allows for an iterative approach where Claude builds what I tell it to. Then I test it and make some tweaks or functional additions. Claude listens to my requests and makes the changes. If something doesn't work it fixes it. I sit back and watch the screen scroll.
Is It Safe?
Nope. Well, really, it depends.
The best way to use this now is if you are already familiar with the language you are coding in. Claude isn't always thinking of security out of the box.
Its also not private. Everything Claude reads and writes it shares with Anthropic, so don't use this if you have sensitive data which cannot be shared.
There is also no guarantee the code will be good. These are still the early days. Humans may be able to code better but AI can code faster. For some people that difference will matter, but for non-developers who just want something that works now, it may not.
The best use case right now is for fairly simple personal tools. I would not expose anything Claude builds to the wider internet unless you know security very well. Nothing that takes personal information or payments. But for things built just for you running locally with LAN only access? This is like magic.