[asc91]

I would like to ask the author, what do they mean when they say 'I spent 2 days and 1M claude code tokens' Did you completely vibe code this? or did you create a spec and asked claude to implement that? How many times did you need to course correct claude when it deviated if it did? How much of your own experience as working in WiFi industry was useful while building this project? And what would be different if you had worked on this project without that experience?

[byb]

Author here: this is actually a "side quest" which I alluded to in the README.adoc and another comment. I've had a few Spotify Car Things (SuperBird) sitting on my desk that I have not touched in a year because I was waiting for the community to come out with a non-tethered approach to reclaiming them. All of the existing continue to either require a host PC, phone, or RPI Zero. I want to deploy these throughout my home and have them usable without my PC or phone nearby - likely with just a Home Assistant dashboard. I spent the first few hours cloning repositories of existing firmware, and examining other approaches, such as Bluetooth low energy with ESPHome Bluetooth proxies, but they ended up being dead ends. I took a compromise that a small wireless dongle would be acceptable. I then began examining the Linux source tree to see if it was possible to flip the USB from client to host and spent a considerable amount of time recompiling BishopDynamics' SuperBird NixOS distribution and building a test dashboard to show the IP networking information, because I knew that once I flipped the device's USB mode, I would lose access to the device. SuperBird took almost four hours to compile on my low end CPU - I let it compile while I hit my five hour Claude Code quota.

https://gitlab.com/baiyibai/superbird-host-mode

I also did not look forward to trying to build in wireless functionality in to NixOS, because it also took 30 minutes to flash the device - I had to lower the transfer block size and add more retries in the flash utility because it kept crashing.

ultimately both the SuperBird work and the RPI pico firmware was about 300 prompts of investigation, research, creating potential approaches, architecting, designing, debugging, testing/validating, and documenting.

Regarding your last two questions... I started out 25 years ago by studying for a CCNA in high school and joined the industry just as first generation 802.11ac devices began hitting the market. From my position, I've worked on every layer from the SoC, PCB, to software. I want to be very careful and say software estimation is very difficult. The main bug (issue #1) I encountered would have been very difficult to diagnose and resolve ten years ago using traditional development workflows.

[asc91]

I'm only starting out in embedded software domain, and the way you describe your 2 day side quest feels like Dumbledore showing some magic trick to Harry.

[byb]

I'm flattered. You asked very good questions. I'd be happy to de-mystify it with a longer conversation. Find me on LinkedIn: 白一百

[wpm]

I've often thought that if I ever did get around to publishing any of the stuff I work on that I would probably also publish any AI usage verbatim. Like, the full logs of all of the chats, at least as a sign of "here was my thought process" and "here's how I convinced the text machine to give me the right text". That somehow feels more honest than "I made this thing, but actually I vibe coded this thing."

[byb]

Author here: I will consider writing a blog about ir, but it will require a lot of work on my part, because as I mentioned, this was a digression/side quest from the main task of hacking my Spotify Car Thing. It might work better as a YouTube video.

[crtified]

Interesting point. The prompts are the source, in a way.