For seven weeks NEURA could hold a conversation. This week I taught it to actually do things with one. That is a bigger jump than it sounds. Answering a question is easy. Carrying out what you asked, reliably, on hardware I own, is what turns a chat window into an assistant.
The core of it is a skills system. There are around forty-four skills right now, and each one is a specific thing NEURA can actually do instead of just describe. Ask it to remember something and it remembers. Ask it to open a tool and the tool opens. Ask it to change how it looks and it changes. Ask it to drive and it drives. None of that leans on the cloud, which mattered to me. NEURA lives on hardware I own, so the actions have to work locally too.
Getting to reliable was the hard part. It is one thing for NEURA to say it did something, another for it to truly happen every time. I spent most of the week in the boring middle, making sure the right skill fires for the right request and that it fails quietly and honestly when it cannot, instead of pretending it worked.
I also split the experience into two clear modes, a robot mode for when NEURA is running the desk robot, and a computer mode for when it is just on the machine. Both keep an easy path back to the face, so I am never more than a step from the thing that feels like NEURA itself.
This is the week it stopped feeling like a program I talk to and started feeling like an assistant that gets up and does the work. Small shift on paper. Big one in how it feels to use.