Claude Fable 5.2 Designed Three Electronic Devices From Scratch Using a Short Prompt

Hardware forgives mistakes far less than software does. A typo in code is fixed in a minute, but a swapped chip pin means a new batch of boards and a week of waiting. That is why a story in which a language model designs an entire device and it powers on the first time is more impressive than yet another benchmark.

What exactly was done

Anthropic ran a series of experiments with Diode Computers. Claude Fable 5.2 designed three devices: a speakerphone with USB-C and Bluetooth, a voice hub for a smart home, and a portable virtual pet.

For each device, the model received a short prompt: an idea plus the main hardware requirements. After that it worked on its own. It wrote the specification, drew the electrical schematics, placed components on the board, designed the enclosure, wrote the firmware and compiled the parts list.

The physical part was done by humans. Diode engineers assembled and launched the devices strictly following the model's instructions. According to Anthropic, all three prototypes worked without any changes to the circuit design and still haven't broken.

In a typical team, these stages are handled by different specialists: a circuit designer, a board designer, a mechanical engineer, a firmware author and a procurement specialist. Here the entire chain went through a single model. That makes the main question consistency between stages. The schematic, board, enclosure and firmware must fit together, and the joints between them are where things most often break.

The mute button and a 25-degree angle

The speakerphone example shows this best. The board could not simply be dropped into the enclosure: the mute switch hit the back wall. Claude proposed mounting the board at an angle of about 25 degrees and added this step to the assembly instructions.

It is a small detail, but a telling one. This is a problem at the intersection of electronics and mechanics: a circuit can be flawless, but if the button doesn't fit in the enclosure, you can't assemble the device. The model kept the electrics, the enclosure geometry and the assembly order in mind at once. And it wrote the solution down so that a person with a soldering iron could repeat it.

What we don't know

There is only one source, a post by Anthropic itself, and no independent verification yet. There are several notable gaps:

• it doesn't say how many iterations and attempts were needed to reach the result;
• it is unknown how much time engineers spent on checking and assembly;
• nothing is reported about certification (mandatory for Bluetooth devices) or mass production;
• "still haven't broken" – the operating period is not specified.

Judging by the description, these are not smartphone-class devices. A speakerphone, a hub and an electronic pet are tasks of a clear, modest scale. There is no data on how the model would cope with a dense multilayer board or strict power consumption requirements.

What to do with this

Hardware startups and product managers. The cycle from idea to working sample may get shorter. It may become possible to test a device hypothesis before hiring a full team. But first consider which stages you are ready to trust to the model and which to check manually.

Engineers. The role is shifting toward review and verification. A human should check the parts list, power circuitry and routing personally until there are reliability statistics. A useful habit from the 25-degree example: require the model to record assumptions and non-standard solutions directly in the instructions.

Anyone following AI. The very framing of the task is telling. It is not a standalone skill like writing a function, but an end-to-end chain from a text description to a physical object on a desk. Agentic models are truly tested on exactly such end-to-end tasks, because there you can't hide an error behind a nice report: the device either turns on or it doesn't.