How my use of AI grew

In January 2026, I was using Claude Code on my Home Assistant Yellow, working on a dashboard for a wall display. It could read the configuration, look through the logs, and help make changes to the actual setup. I could explain what I wanted, have it investigate, and work through the result with it.

I wanted to see what else it could do. Soon I was asking whether it could help on my other devices, too.

Within days, I was working with Claude Code across several machines. Then came shared storage and ways to hand work between assistants. If one had already figured something out, I wanted the next one to have that information. Having to explain the whole project again was a problem worth solving in its own right.

What excited me was how far I could follow an idea. I could work through the options, have something built, try it, and come back with a much more specific request. AI tools wrote the code. My part was deciding what I wanted the thing to do, understanding the choices, and checking what came back.

An early gaming project involved converting and transferring files, verifying the outputs, and giving the process an animated progress display. I care about that last part. I want things to work, and I want to enjoy using them. Being able to see what’s happening is useful; giving that experience some personality makes it more interesting to spend time with.

There were also occasions when a lot of work developed around an assumption that should have been checked first. One home-automation project had an elaborate detector that expected sensor readings much more frequently than they actually arrived. Until that was corrected, we couldn’t sensibly judge the detector.

Once an assistant starts building on an assumption, it can get a long way. I wanted better ways to catch those problems before more work depended on them.

In February, I asked for a constitution for my AI agents. I wanted rules about permission, evidence, safety, and what it meant to say a job was finished. I asked for checks and balances, and we ended up with branches of government and a separate legal code.

An audit later that month found that much of this machinery wasn’t connected or working as intended. We cut it down substantially, keeping a smaller core that included checks capable of blocking dangerous commands before they ran.

A few months later, an inventory drifted out of date without the existing audit checks catching it. I asked for another examination from the beginning, including whether we should scrap what we had. My own design had to be open to revision, too.

The rules evolved to cover consent, evidence, recovery, records, and confidentiality. Getting them enforced consistently across different tools is still ongoing work.

Meanwhile, I was finding different uses for different assistants: exploring an idea with one, having another implement it, and asking a third to challenge the result. I’ve carried plenty of those handoffs between sessions myself. I want cloud and local agents to coordinate that work, use the tool suited to the job, and keep track of decisions we’ve already made.

Fleet Orchestrator is where I want that experience to come together. A project I saw on Reddit represented AI workers as little characters. I loved the idea. I wanted original characters of my own, which I call glyphs, moving and interacting in ways that help me understand what the team is doing.

I want to replace all those separate windows with one place to see the work. If a different model would be better for the next part of a job, I want to switch the worker over and carry on. Animated glyphs, persistent sessions, and model handoffs are already parts of the project. Making the whole experience work smoothly is still the ambition.

That’s where I want to get: open one window, see what the team is up to, and have enough understanding to make a useful decision about what happens next. And yes, I want it to feel a little magical.

How I work with AI tools →