Arguably you could just translate the STM32 LL HAL though, and I'm sure they trained on it, cos it sometimes regurgitates that. So I wonder how much of this is actually using the HTML and how much is just in it's training data. Maybe giving it access to the LL headers would make it even better?

Hm, it's not great at logic problems like finding out which registers are available on which devices. But at least it can write helper tools that are *almost* correct (see the CRGR3). I need to clean this up and scale it to all STM32 devices and put it online.

To be clear: I still have to review and fix the code for *hours* because the coding agents simply do not understand the logic part of generating C++ HAL code using Jinja2 templates. I would have to provide them tooling to query the differences and find short filters, but if I have these tools, I might as well write the code myself again. So we are again mainly limited by tooling for hard problems.

However, what this thing is really good at: rage baiting me into reviewing the generated code. It unblocks me by just *starting* the task, even if it's shitty. I often am unmotivated to refactor because of the mind-numbing amount of just… pushing code around. The coding agents make that much less painful, even if it's imperfect. I just kinda wish it were available locally.

Another takeaway from this experiment is that none of these datasheet AI companies have any moat. I can dump in my HTML versions of the PDF reference manual and ask it questions, and the answers are very, very accurate. There were several parts that I didn't fully understand in the RM, and the LLM just gave me code that I was able to walk through with the RM for better understanding. All these companies are just one prompt extraction away from bankruptcy.