Comment by mshafir

It resonated with me too, and even as a once fairly prolific coder that now uses agents to do all the coding, I feel like this is all relevant and does matter. The agent does whatever you ask and provides thin judgements based on what it thinks is your motivation and context, even feeding it the context it will often make bad generic choices. I think this is the part hard to optimize because it's not a verifiable reward, it's the feel of the software. I think Herdr is a great example, I've tried several dozen terminal multiplexers / session aggregators and they are all vibed and different, but this one felt so much better to use because of its choices and actually increased the number of parallel sessions I was able to keep going. My own projects I find I have to stress every detail, that is what makes the software good now is really diving into it, the experience feels like just as much of a craft as coding was and 'taste' captures the concept for me.