An AI Designed a Working Circuit Board From One Plain-English Prompt

A hobbyist set out to design a printed circuit board without learning any PCB tools, relying entirely on Anthropic’s models to do the work. An earlier attempt with Claude Opus 4.8 flopped—it botched component orientation and routing—but the newer Fable 5 model succeeded under two self-imposed rules: no manual edits or verification, and every pre-manufacturing problem had to be solved by the AI. From a short description of an RP2350-based development board driving an E-ink display, the model worked autonomously for a few hours and produced a 4-layer, ~32×37mm design with an RP2350A MCU, a 1.54-inch 200×200 E-ink panel, and 8MB of QSPI flash, costing about 26€ per board.

The process was far from clean. The initial layout carried 65 design-rule-check errors, and the model picked wrong footprints for two parts—the SPI flash chip was too big for its pads and a boost-converter transistor too small—forcing a substitution to a P25Q64SH after back-and-forth. Autorouting was the weakest link: the open-source Freerouting tool plateaued after 17 passes and left 49 connections unmade, so the AI hand-routed the rest. The author later found the KiCadRoutingTools project finished the same job in 1.25 seconds. Despite this, five assembled boards ordered through JLCPCB for 130€ arrived working, powered up without shorts, and ran proof-of-concept apps.

The piece is as much a meditation on deskilling as a build log. The author is impressed that someone with zero PCB knowledge can describe a board in English and receive functional hardware, but admits the experience drained the satisfaction that once came from struggling to learn. The takeaway for a technical audience: AI is now capable enough to carry a real hardware project from prompt to manufacturable output, yet still stumbles on concrete details—package selection, footprint matching, routing—that require either better tooling or a human who knows what to check.