JerkyTreatsI don’t read the code, but I’m shipping higher quality, more ambitious software.
Generating code is cheap now, measured in time-to-completion, against the human-only prior.
As a result, my time has shifted to either side of code generation.
On one side, you have systems design, acceptance criteria, expected outcomes, software architecture. I put an intense amount of focus on the shape of what I want to build. It’s reducing complex problems into ownership domains, clearly defining the boundaries of those domains, and defining a system that represents those domains and gracefully connects inputs to outcomes.
On the other is validation. This is testing, verifying, refining, refactoring. It’s capturing misses into a continuously expanding test harness. Unit tests to integration tests to local stack to production stack. OTel instrumentation for agent readable runtime performance.
And oh my God is it ever fun!
Manually writing code was such a slog. You had to train yourself to endure long painful sessions of it not working, and trick yourself into believing it was worth it when you finally figured it out. Sometimes it was. Sometimes it was literally just step one.
I won’t deny that there was valuable learning in the pain. But, you can’t dismiss the valuable learning in just asking the AI how thing works. A rubber duck that talks back is worth the cost of admission alone.
But now? Holy shit I’m building things I’ve been dreaming about all my life. Stupid fascinations like Hierarchical Task Networks with dynamic knowledge graphs. Polygenic Risk Calculators connecting to genetic research databases.
Ambitious stuff, for I am no geneticist.