- AI
- Workflow
- Opinion
AI-first is a workflow, not a feature
Everyone is bolting a chatbot onto their product and calling it AI-first. The real shift is in how software gets built — and what it does to code review, testing, and taste.
Every second landing page this year says "AI-powered." Most of them mean a textarea that forwards to a model API. That is a feature — and often a good one — but it is not what AI-first means to me as an engineer.
AI-first is a workflow. It changes how the software gets made, before it changes what the software does.
The loop has inverted
The old loop: think, type, run, debug. The new loop: specify, review, steer. I spend less time producing code and far more time reading it — auditing a diff an agent produced against an intent I wrote three minutes ago.
That inversion has consequences nobody puts on the landing page:
- Review is the bottleneck now. Generating a 400-line change takes a minute. Understanding it still takes as long as it ever did. Teams that merge on vibes will discover this the expensive way.
- Tests stopped being optional. When a human writes code, the tests verify the code. When an agent writes code, the tests are the spec. A weak suite means the agent optimizes for "compiles" instead of "correct."
- Taste compounds. The model happily produces five workable designs. The engineer's job is knowing which one you will not regret in six months.
The model writes the code. You are still responsible for the system.
What this looks like in practice
On my own projects, the workflow has settled into three habits:
- Write the constraint, not the code. A precise paragraph about invariants and edge cases beats a vague prompt plus four correction rounds.
- Keep the diff smaller than your attention. If I cannot hold the whole change in my head, I split the task — not because the model can't do more, but because I can't review more.
- Automate the boring verification. Typecheck, lint, tests, build — every agent turn ends by running them. Machines checking machines, so my review time goes to design instead of typos.
The uncomfortable part
Some of my hard-won skills depreciated. Memorizing an API surface is worth little when the model knows every API surface. What appreciated instead: systems thinking, data modeling, knowing what "correct" even means for a given feature. The mathematics degree turned out to be a better long-term asset than the jQuery years.