Pairing with AI: how LLMs changed the way I build
From FTP uploads to AI pair programmers in one career. What large language models are actually for, and the team rules that make them safe.
I was skeptical for a long time. I watched juniors paste half our codebase into a chat window and accept whatever came back, and I thought: this is how quality dies. Two years on, I use an AI assistant almost every working day — and I've changed my mind about what it's actually for.
It's not an autocomplete. It's a rubber duck that talks back
The least valuable thing an LLM does for me is write code. The most valuable thing is compressing the distance between a question and a starting point. "How would you structure a multi-tenant billing migration?" used to be an afternoon of reading. Now it's ten minutes to a draft I can argue with. The arguing is the work — it always was.
The model writes the first draft. Your judgment is still the product.
What we changed at the team level
- Review standards went up, not down. Generated code reads confident even when it's wrong. We review it like code from a bright intern with no context — because that's what it is.
- Boilerplate died. CRUD scaffolding, test skeletons, migration files, one-off scripts — nobody types these by hand anymore, and nobody misses it.
- Juniors need guardrails. The danger isn't that they use AI; it's that they skip the struggle that builds judgment. We pair AI use with "explain this line back to me" reviews.
The long view
I started my career hand-writing SQL joins in PHP files uploaded over FTP. Every wave since — frameworks, containers, cloud, and now this — moved the effort up a level of abstraction and made judgment matter more, not less. This wave is no different. The engineers who thrive won't be the ones who type fastest; they'll be the ones who know what to ask for and how to tell when the answer is wrong.