I’m fine if 100% of my code is actually from Claude
Most of the time, I like to use Claude Code to make targeted changes along with the tests. I try the feature, notice issues with the change, and clarify with better tests along the way.
Occasionally I give it a more ambitious task, but the variance increases, and much depends on whether I’ve given the right guardrails. Like recently I upgraded a pandas plugin to Pandas 3.0, but it had heeaaavy amounts of testing.
In both cases my software brain is still fully on and I care about the same things I used to care about. I’m just not wasting time on dumb human mistakes. Why would I? I have a predictive model of language next to me
I do still write a bit of code (like in a Jupyter notebook). Sure it’s a tad rusty, but it’s perfectly fine. I don’t yet feel completely decayed in my ability to appreciate code.
In the end we have always had to go up/down levels of abstraction and not get too attached to one tool. I see using an LLM the same, you can use it in tight, low variance changes. Or go big with bigger swings. Sometimes you vibe code a whole app to test PMF and create a demo. Sometimes you’re really fine tuning 10 lines of code for performance and correctness.