AI-Assisted Coding

AI-assisted development — the thing people call vibe coding — shifts where your time goes. Less typing, far more specifying and reviewing. The engineers who get the most from it are the ones who treat it as delegation rather than autocomplete.

Specify before you generate. A vague request produces plausible code that solves an adjacent problem. Say what the input and output are, what the constraints are, what should happen on failure, and what existing patterns to follow. Most disappointing output traces back to a prompt that did not actually contain the requirements.

You own everything you merge. The author of record is you. Code you cannot explain in review is code you should not have merged, and "the model wrote it" is not a defence anyone accepts — nor should it be.

Work in reviewable increments. A 40-line change you read properly beats an 800-line change you skim. The failure mode is accepting a large plausible diff, discovering a problem three features later, and no longer knowing which assumptions are load-bearing.

Be most careful where you know least. Generated code in a language or framework you know well is easy to check. Generated code in one you do not is where confident errors survive review — and exactly where the temptation to accept it is highest.

Tests are not free verification. A model that misunderstood the requirement will write tests that confirm its misunderstanding. Tests generated from the same wrong premise pass happily. Write or at least specify the important assertions yourself.

Watch architectural drift. Each local suggestion can be reasonable while the sum is incoherent — three ways to fetch data, two error conventions, duplicated helpers. Nothing in the loop is holding the whole design in mind unless you are.

Keep a fast feedback loop. Types, linting, and tests running continuously catch a large share of issues immediately, and they matter more, not less, when code arrives faster than you can read it.