What they did
Matt Inglot, a software consultant who builds electronics as a hobby, runs a yearly haunted circus and wanted one standard board to trigger props, play audio, drive LED strips and be tuned over Wi-Fi. He used Codex, first with GPT Sol 5.6 and later GPT 6 Astra, to turn a wish list and a critique of his first board into a numeric spec. The agent then searched JLCPCB stock through its API, read datasheets, picked the regulators, amp, DAC and ESP32 module, wrote per-subsystem design notes, ran reviews and simulations, edited the KiCad schematic files directly and later routed the traces. He placed the parts by hand. Each time he coached it through a hard step, he had it write a skill so the lesson stuck.
How it went
Ten boards cost about 680 USD delivered, plus customs, and every major feature worked on the first spin. The first schematics were electrically mostly right but badly laid out; they improved once the agent exported KiCad renders to look at its own work. First routing attempts were poor, and after rounds of feedback and its own checking scripts the result was, in his words, roughly as good as his own. The power design was reworked several times because the agent over-engineered and misread loose spec ranges. A long ngspice simulation produced nothing useful until it was scoped down.
Worth knowing
This took two months of close supervision and prior PCB experience, not a one-shot prompt. Be precise with spec wording and budget, since the model padded tolerances and spent freely by default. A commenter who tried similar work found the agent good for footprints, pin mapping and checks but slower than doing schematics and layout by hand. He also had Claude design a 3D-printed enclosure from the KiCad export.