I’ve got more than ten half-finished maker projects sitting around my workshop, and the ideas keep coming faster than I can clear the bench. So I’m running an experiment: could AI help me get projects through to completion faster?
I starting with Arduino work because it felt like the best fit. I’m not a natural programmer. I can find code, make it work, and tweak it, but I lose time when I have to keep patching things after I’ve changed them. AI should be good at cutting that part down.
Why this Raspberry Pi mattered
The Pi I used belonged to my dad. He used it for his 3D printing OctoPrint setups. Bringing it back into my workshop as part of my day-to-day workflow felt personal, not just practical.
The first part was all on me. AI cannot hold a screwdriver. I had to flash the drive, deal with boot trouble, and get the Pi running before I handed anything over to the agent.
Getting the Pi ready
I used the Raspberry Pi Imager and set up a USB boot drive with the Raspberry Pi Minimal image. At first, the Pi would not boot from USB. I tried checking my DHCP logs, but there was nothing there, so I needed to see what the screen was saying.
I did not even have the right mini HDMI cable at first. Then I spotted a micro HDMI cable and used that to connect the Pi to a monitor.
Once I could see the screen, I found the real issue: the bootloader needed updating. It was still trying to boot from the SD card instead of USB. I reflashed the bootloader using the Raspberry Pi utility meant for USB boot, and after that the Pi started up from the USB drive.
Handing it over to Frank
My AI agent is called Frank. It runs on the Hermes agent harness, and I used GPT-5.4 mini because it is token efficient.
The first job was to set up mDNS so I could reach the Pi by hostname on my network. That way I would not need to keep chasing changing IP addresses.
The second job was to install the Universal Embedded Workbench, an open source Arduino and ESP32 workbench project on GitHub by SensorsIot. Andreas Spiess has also been involved in showing how this kind of setup can be used with an AI agent.
My first prompt was too vague. Frank tried to install the workbench locally on my machine instead of on the Raspberry Pi. I spotted the mistake from an error message partway through. I corrected the prompt and told it clearly to install on the Pi, not locally. After that, it worked.
Testing the setup
Once the workbench was installed, I gave Frank a simple test: blink an LED with a Morse code message spelling SOS.
That worked too. I also had Frank save the code to a shared drive so I could inspect it myself afterwards.
One thing I noticed is that the more text the AI produces, the more output tokens it uses. If I want detailed comments in the code, that helps with understanding, but it also adds to the cost. For my use, I want the balance between useful output and wasted tokens.
What I learned
This is only the start. The main lesson for me was simple: AI agents are useful, but only when I give them a clear job and the right guard rails. If my prompt is vague, they can go off in the wrong direction.
Even so, the result was promising. I now have a Raspberry Pi set up as a wireless Arduino workbench, and that opens the door to using AI to help with more projects in my shop.
I still want to stay in charge of the project direction. I decide what the job is, what it connects to, and what the pins need to do. The AI can do the coding while I keep moving the project forward.
Next, I want to see where else this could help me, whether that is 3D printing, laser cutting, or CNC work.