AI Designs a PCB from Scratch: Claude Builds a Working Circuit Board That Lit Up First Try
Claude autonomously designed a working PCB with KiCad MCP — and it lit up successfully on the first try.
A hardware enthusiast handed a years-delayed PCB project entirely to Claude, with a strict rule of no manual intervention. Using the KiCad MCP toolchain, Claude worked autonomously for hours to design a Raspberry Pi Pico 2350 development board with an e-ink display. Component selection errors surfaced during fabrication but were fixed through conversation. The five boards ordered for €130 showed no short circuits on power-up, and the firmware ran perfectly first try — proving AI-assisted PCB design has crossed the usable threshold.
From Idea to Hardware: A Bold AI Hardware Design Experiment
For many hardware enthusiasts, turning a circuit idea into a real PCB (printed circuit board) — and then writing firmware to drive all its features — is an exciting but intimidating process. One Reddit user shared his personal experience: a PCB design he'd been putting off for years, finally handed entirely to the AI model Claude, with the finished board actually working.
This wasn't just a casual experiment — it was a genuine test of AI's capabilities at the frontier of professional hardware engineering. The author set two nearly uncompromising rules: no manual modifications or verification, and every problem encountered before manufacturing had to be solved by the AI. In his own words, this meant he had to "trust Claude with his wallet."
Early Failures: Older AI Models Couldn't Handle PCB Design
The road wasn't smooth from the start. When the author first tried using earlier models to "vibe-generate" circuit boards, the results were deeply disappointing. Those models had no concept of component orientation and couldn't produce reasonable routing. He briefly concluded that AI-assisted PCB design simply wasn't ready for real use.
This highlights just how challenging PCB design is for AI — it requires not only understanding the logic of electrical connections, but also handling spatial layout, component footprints, and physical constraints across multiple dimensions. That's far more complex than generating a snippet of code.
The Turning Point: A Natural Language Prompt Generates a Complete PCB
The real breakthrough came on a Thursday evening. With about 10 hours left before Claude's weekly usage quota reset, the author decided to try again with a newer model. He positioned himself purely as an "end customer," staying out of the design process entirely and describing his requirements in a short, plain-language prompt.
One Prompt, One Complete Development Board
His core prompt went roughly like this:
Design a PCB featuring a Raspberry Pi Pico 2350, four buttons (two on each side, using Watchy-style buttons or equivalents), and a GDEY0154D67-FL04 e-ink display. The PCB should match the display's dimensions. The Pico drives the screen; the buttons control the Pico. Also include an I2C interface and some GPIO breakout pins. Use KiCad MCP, and let me know when the schematic and layout are done.
That single paragraph described a development board based on the RPI 2350, capable of driving an e-ink display. After working autonomously for several hours, Claude delivered a complete design. The key enabler here was "KiCad MCP" — through the Model Context Protocol, Claude could directly operate the professional EDA tool KiCad, translating natural language requirements into real schematic and PCB layout files.
Reality Check: Does an AI-Designed PCB Actually Work?
Of course, AI-designed PCBs aren't without flaws. When uploading the files to JLCPCB for fabrication, the author discovered that Claude had selected the wrong parts for two critical components: an SPI Flash storage chip and a transistor in the e-ink display driver circuit. On top of that, some of the recommended components were out of stock on LCSC.
A Few Rounds of Conversation Fixed All Component Issues
Interestingly, these problems didn't derail the experiment. After a few back-and-forth exchanges, Claude corrected all the component selection and stock issues. The author ultimately spent €130 to have 5 boards fabricated, and purchased the e-ink display separately from a local store. True to his word, he ordered "with his eyes closed" — no manual intervention at any point.
Power-On Moment: No Short Circuits, Lit Up First Try
The boards arrived a few days later. Drawing on past experience burning out several Framework laptop USB modules, the author first checked for a short between 3.3V and ground with a multimeter — even though his colleague suggested just "YOLO-ing" it and plugging it in. The result was a pleasant surprise: no short circuit. When connected to a computer, the board was recognized immediately and ready to use.
He then wrote the driver software, which ran perfectly on the board, with the e-ink display lighting up and functioning correctly. From PCB design to fabrication to firmware, the entire chain was AI-led — and it worked on the first attempt.
The Rapid Evolution of AI Hardware Design Tooling
The author mentioned one more detail at the end: after placing the order, someone recommended another auto-routing tool called KiCadRoutingTools, which is apparently significantly faster and more efficient. He plans to use it in future projects. This suggests that the tooling ecosystem around AI-assisted PCB design is maturing quickly — dedicated auto-routing tools can further compensate for large language models' shortcomings when it comes to physical layout.
Excitement Mixed with Reflection
What's worth noting is the author's complex feelings about the outcome. He was very satisfied with the result, but candidly admitted he "didn't feel much of a sense of accomplishment." He still wants to properly learn real PCB design, and in the meantime, plans to keep using AI to generate boards for hobby projects.
This tension reflects a broader dilemma brought on by AI tools: when AI can do professional work for us, have we truly gained a new capability, or have we simply bypassed knowledge we should have built up ourselves? For lowering the barrier to entry in hardware making, this is undeniably a huge step forward — but for developing deep engineering competency, it raises new questions worth thinking about.
AI PCB Design Has Crossed the "Usable" Threshold
The biggest takeaway from this case is what it proves: with the right model and toolchain (Claude + KiCad MCP + JLCPCB professional fabrication), AI can now end-to-end design a development board that is structurally sound, manufacturable, and working on first power-up. There are still rough edges — wrong component selections, manual stock verification needed — but these can be resolved quickly through conversation.
From earlier models being "completely unusable" to now "trusting it with your wallet" and successfully shipping, the pace of AI progress in PCB design and hardware engineering is remarkable. For hardware enthusiasts everywhere, the era of rapidly turning ideas into physical products may truly be arriving.
Related articles
Catalyst: A Vision for an Enzyme-Like Testing Framework for AI Agents
A developer shared Catalyst on Reddit, an Enzyme-inspired framework for AI Agents, exploring why agents need observable, testable dev tools and the design philosophy behind them.
The Real Capability of AI Coding Agents: Best Models Complete Only 35% of Feature Development Tasks
The 'Agents on Rails' benchmark finds top AI models complete only 35% of feature development tasks. What this means for coding agents and developer teams.
How to Prevent Duplicate Refunds After an AI Agent Crashes: CellaFlow's Durable Execution Approach
How can AI agents avoid duplicate refunds after a crash without deadlocking workflows? CellaFlow uses durable execution, shared work identity, leases, and fencing to solve safety and liveness in multi-agent systems.