I see the value in resurrecting old projects for fun. But I still struggle to use it for anything other than research and small changes.
For example, I have a few personal projects I’m working on. If I create a spec, and allow agents to iterate, I still need to validate; validate not just tests, but my understanding of what’s going on in the system. And when I try to validate, I get overburdened from trying to understand a codebase I didn’t write.
I’m not saying I fully understand everything I’ve ever written, but that crucial system understanding comes from writing it (in my opinion of course). This isn’t even mentioning the learning that comes from writing.
Again, I use it for researching, small changes, and even tests, but I struggle to see the benefit of letting it write the code for me, unless I want to atrophy my skills, knowledge, and understanding (which I’ve seen in my coworkers).
“I’ve dialed back Claude’s enthusiasm for explaining everything. Instead, I can bounce ideas and problems off it. Often it comes back with something that triggers a brand new direction.”
That sounds interesting. How did you do it, and what does it look like in practice?
I self-review every line of AI-written code that goes into work projects, and Claude is doing a pretty good job these days of writing that code in at least as clean a way as I would have done it myself (and where it doesn’t, it gets fixed).
Claude, and Codex, have allowed me to successfully ship projects in domains where my understanding is lower, in languages where my experience is limited. I could never have imagined building an assembler. Or a cycle-accurate emulator, much less several of them. And my lack of experience with Rust, and Swift, meant that even bootstrapping such projects was challenging in the past.
Obviously execution is the hard part, but now I can at least prototype an idea and get working software - it may not be perfect, there may even be serious architectural issues, but I can validate the concept before doing it for real.
I see the value in resurrecting old projects for fun. But I still struggle to use it for anything other than research and small changes.
For example, I have a few personal projects I’m working on. If I create a spec, and allow agents to iterate, I still need to validate; validate not just tests, but my understanding of what’s going on in the system. And when I try to validate, I get overburdened from trying to understand a codebase I didn’t write.
I’m not saying I fully understand everything I’ve ever written, but that crucial system understanding comes from writing it (in my opinion of course). This isn’t even mentioning the learning that comes from writing.
Again, I use it for researching, small changes, and even tests, but I struggle to see the benefit of letting it write the code for me, unless I want to atrophy my skills, knowledge, and understanding (which I’ve seen in my coworkers).
Would love your insight into this.
“I’ve dialed back Claude’s enthusiasm for explaining everything. Instead, I can bounce ideas and problems off it. Often it comes back with something that triggers a brand new direction.”
That sounds interesting. How did you do it, and what does it look like in practice?
I self-review every line of AI-written code that goes into work projects, and Claude is doing a pretty good job these days of writing that code in at least as clean a way as I would have done it myself (and where it doesn’t, it gets fixed).
Claude, and Codex, have allowed me to successfully ship projects in domains where my understanding is lower, in languages where my experience is limited. I could never have imagined building an assembler. Or a cycle-accurate emulator, much less several of them. And my lack of experience with Rust, and Swift, meant that even bootstrapping such projects was challenging in the past.
Obviously execution is the hard part, but now I can at least prototype an idea and get working software - it may not be perfect, there may even be serious architectural issues, but I can validate the concept before doing it for real.