I use Claude Code for intensive robotics programming. I used it recently to advise on KiCad schematic and PCB board layout. I was surprised when Claude gave me recommendations and even offered to make the changes automatically. When I let it proceed it modified both schematic, checked with LCSC for available components and placed the components on the PCB. I did not ask about routing the traces because I wanted to test Quilter for that. So bottom line, what I learnt was that I do not need to be fully conversant with KiCad since Claude can be given full access to sketch out the schematic and layout PCB components with footprints that I can then review. No real need for hours of manual KiCad work😊

Well-thought-out review - thank you. I like the way you think about this. Excellent!

I am interested to see if you use Planning mode to plan out the circuit first with a couple of iterations and plan grilling + research to see how well it can then design a circuit efficiently enough with the correct parts.

An interesting test would be to see if it can draw up a simple Eurorack module. Something like a Voltage Controlled Oscillator.

Check out matt pocock's skills if you haven't already.

For generating base schematics, coming up with e.g. interlocking power sequencing (as in "once USB is connected, the boost converter can't provide any power"), the LLMs nowadays are remarkably good. Don't yet let it design PCBs and triple check the footprints and components.

I agree. Its a great video.
Not everyone blindly delegates important stuff to AI - like I didn't know CFD, FEM until AI made it accessible. Now electronics is now Accessible.
Just watching the design process is great.
I noticed you didn't use a harness to make it less prone to hallucinations, or set up a components and spec database, and its a one off simple prompt not the managed iterative prompts.
Then there is the knowledge you get from failure and working that easily becomes new SOP.

I know AI would destroy few jobs and is based on stolen knowledge, but damn if I live long enough to make complete electronic devices by myself without knowing anything about electronics, that would be amazing
Anyway thanks for interesting video ;)

Darn. I'm on the 20 dollar plan too 😂 Pretty cool. We'll see if and how fast it improves.

How much you want to bet the distance between the headers isn't a multiple of 100 mils?  Try to plug this into a dev board and the pins won't line up horizontally. Honestly looks kinda wide too, looks almost like once you plug it in, there won't be space for jumper wires beside it.

I remember 4 years ago with the Original release of ChatGPT, my coworker said that AI would get nowhere, and it would be incapable of doing CAD. But here we are 4 years later. You are probably not using frontier models also. I hope AI hits a limit otherwise at this rate. We will be doomed in 2-3 years.

The MCP was totally unnecessary, just give it the folder, it can do all these directly on kicad.

Its wierd theyre spending so much time on this. Are there really that many people who are going to know what the board needs to do, conponent selection ect but not know how to layout? Or worse it does everything for you until you get the board back and somethings wrong. You would have no idea where to even start. I feel like theres better things to develop than eda for people to lazy to take a weekend to learn something

This is the worst it’s gonna be…

Ok, it is time to regret having shared kicad projects on github for years. May open AI burn and Anthropic shares be given back to ftx victims. Already even before llms, people said 'you can be replaced at anytime' even if it was bluf. Now, no one will thank the open source comunity for the data, just a 'now you lose your job'. Contrary to code where it speeds up the work, I would not use this stuff. It is like reviewing work from an incompetent. Maybe someday it will be reinforced with emf simulations. But in a human or social terms it is shit. Hey by the way, you do not need to put a transistor on a micro controller pin to run a relay...

The tool doesn't need to be improved, YOUR usage of it does. You're asking it to do a lot at the same time (not to mention reading your mind), Circuit design, component selection AND optimization from an operating perspective AND price point with an ambiguous prompt in ONE shot. How much did u spend on tokens for this output? compare it to a contractor given the same prompt (no specifics), would even give that to a contractor?? You would normally give it more of a complete design spec. (you're using a 20$/month subscription??). "i could have maybe used the tool better" - 100% man... complete hack review.. imagine resting a car's performance handling and not considering what road you're using.. "ahh let's just throw it on the gravel" ... or weather conditions... "ahhh it's a bit rainy, whatever"... then saying "yeah i could tested under proper and varied road conditions and weather ...but hey.. whatever" .. perfect example of the junk on youtube. Here's an example of a proper youtube test/tutorial in 2026 - https://www.youtube.com/watch?v=M5KOgtk9VfI

Have proper skills and rules for it, if you let it one shot it most probably fails or does a poor job, it does the same thing with software, but when used in an already established project it can basically do what you ask of it.