The Tools Changed, So I Did Too
A year ago, AI was something I asked questions. Today it's something I work alongside. I didn't plan that shift. It happened one small concession at a time, and by the time I looked up, the way I do my job had quietly rewired itself.
I want to be honest about how that actually went, because the marketing version ("AI changes everything overnight") isn't what I lived.
The model is not the product
For a long time I judged each new release by the model alone. Smarter, faster, bigger context. That was the wrong thing to watch.
What actually moved the needle was the harness: the terminal, the tools, the ability to run a command, read the output, search the docs, and try again without me in the loop for every step. The same model is a clever assistant in a chat box and a genuine collaborator in a good harness. The intelligence was necessary. The scaffolding around it is what made it useful.
So I stopped chasing benchmarks and started investing in the setup. The plumbing is the work now.
I'm not vibe coding, I'm supervising
The honest description of my workflow isn't "the AI writes my code." It's closer to working with a fast, eager teammate who needs clear direction and a careful review. I hand off a well-scoped task. It goes off and does the work. I read the result the way I'd read a junior engineer's pull request: looking for the thing that's subtly wrong, the assumption that won't hold, the shortcut that'll cost us later.
Most of the value isn't in the typing it saves. It's in the drudgery it absorbs so I can spend my attention on the decisions that actually matter. That trade is the whole game.
Every new model resets the floor
Here's the part that's genuinely hard, and that nobody warns you about. The thing you learned last quarter expires.
A new model lands and the prompt that used to need three careful paragraphs now works in one line. The workaround you were proud of is suddenly dead weight. The task you'd written off as "too unreliable to delegate" is now boringly routine. Half of getting good at this is learning the tools. The other half is being willing to throw away what you learned the moment the floor moves.
I've made peace with that. I treat my setup as a draft, not a monument. I re-test my assumptions when a new model ships instead of defending the way I did it last time.
What I tell my team
Don't wait for the perfect moment to start. There isn't one. The skill of working with these tools is the kind you can only learn by doing the work badly for a while first. Tinker. Hand off something real. Watch where it fails. Adjust.
And keep your standards. The tools got dramatically better at producing output. They did not get better at deciding what's worth producing, or at telling you when something is quietly wrong. That judgment is still ours. It might be the only part that stays ours.
The tools changed. So I did too. I expect to keep changing, because they certainly will.