Well, it's finally happened: my daily work routine revolves around prompting Claude. I can't say that I'm sad about this development actually, even though I dislike the general invasion of AI into all aspects of computing. What I do want to reflect on right now is two things: the nature of my work and the nature of my language in communicating with Claude.

When I was younger, I did not strive to become a software engineer; it is simply what happened. For me building software has never been about the process, though I can very much appreciate elegant code and want my output to be as elegant as possible. The software is a means to an end: it must serve a purpose. Thus, the craft of writing code is something I perfected for its own sake, more as a sidequest rather than what I was after in the first place.

Before I understood what "software engineering" was, I think what I really wanted to be was a "software architect". I have always had ideas of things I wanted to build, and ideas from different disciplines I wanted to combine. In the same way that writing good code was something I had to pick up to achieve my goals, learning computational geometry and computational electromagnetism were also simply means to ends.

Claude has completely changed the nature of my work in that I now do not need to have Ph.D level understanding of all these subfields I am trying to combine. I have enough understanding (perhaps at an undergraduate or master's level) to be able to make the informed decisions and suggestions to push Claude into a fruitful direction. As a result, I am simultaneously much more productive as well as much more mentally exhausted at the end of each day. I can see fairly clearly that this is not a sustainable situation. My current sprint is working through a backlog--almost two decades long now--of projects that I not only did not have time to pursue, but which I lacked the background to even start. If there is one thing I would advise students today: learn as much as you can about a very wide variety of topics; be a polymath. I'm glad I took classes on geometric modeling, drafting, and general relativity in college. They gave me the vocabulary and the introduction to concepts that I have borrowed to produce novel products of their combinations.

Claude has stunned me enough times in the last months that a single blog post certainly will not suffice to cover it; I have had a few Jacobian conjecture moments of my own. People have remarked on a few occasions on my relative success in using Claude and getting good output. I think the crux of it is that you really do need to be a domain expert (at least enough of one) to not only sound like you know what you're talking about, but also to know enough to critique and suggest new avenues. As I had put it to a coworker: it is like you're a Ph.D advisor guiding a student, except Claude obviously knows more about everything, so it's more like the student guiding the advisor. All of this is to say, you cannot be completely ignorant and expect good results. You need to be a "jack of all trades", and it's okay if you're "master of none".

On a slightly related note, I want to talk about language and conversing with an LLM. As some have noted, it helps to be polite and professional. It's hardly more effort, and makes the chat log more presentable for sharing with others. That said, I think there is also a degree of mirroring. For me, I have a very strong mirror reflex; when I speak with my neighbors around town, a country drawl comes out. On work meetings, I speak with a fairly standard bay area dialect. With Claude, I find my self using Claude-isms. Words like "decisive", "load-bearing", and "informs" get used much more frequently than I normally would. Does it help? I think so, maybe slightly. I think the more significant aspect for me is that I have improved my understanding of when writing feels like it came from an LLM. A HN post got lambasted for being LLM generated, but I did not feel that it was. I'm not suggesting I'm even better than average at detecting this, but I think it helps me keep my voice my own by being able to both consciously and unconsciously switch my mode of language synthesis when interacting with Claude.

I promised followup posts about LLM-produced projects, but those will take quite a bit longer to write.