'Claude is great at generating boilerplate, why do you still hand-write it.' Someone said this to me last week and it stuck, but not in the way they meant.
I've been using Claude 4.7 and Sonnet 4.6 for about six months now, mostly for prompt engineering work. The thing is, they're genuinely good at scaffolding. I give them a loose spec, they spit out a structure that's 70% usable. I used to spend two hours on that. Now it's fifteen minutes.
But (and I think this is what the comment was really about), I've noticed I stop thinking once the AI gives me something that looks right. Not always. Just often enough that I catch myself doing it. I'll get a generated prompt, glance at it, add some context tweaks, and ship it without actually running it against the use case in my head first. The AI made it so easy to get to "done" that I skipped the part where I actually know why each piece is there.
Sonnet especially is fast enough that it feels like cheating. You write a bad request and get back something coherent anyway, which sounds like a win until you realize the AI hallucinated around your unclear thinking instead of forcing you to clarify it.
That said, I'm not going back to hand-writing everything. The speed is real. It's just that the help comes with a trap door: you can move faster without learning anything, and you won't notice until something breaks. My first paid product just sold, actually, and now I'm wondering how much of what I built I actually understand versus just assembled. Probably regretting shipping it in a week.