How AI Is Changing the Way We Write Software

AI pair programmers have gone from novelty to daily tool in a couple of years. They are genuinely useful — and genuinely easy to misuse. Here is how I think about them.

What they are great at

• Boilerplate: configs, type definitions, test scaffolding, data transforms.
• Explaining unfamiliar code or error messages in plain language.
• Exploring alternatives quickly before committing to a design.

Where they fall short

Models are confident even when they are wrong. They do not know your product, your users or the constraint your teammate mentioned in standup. Treat every suggestion as a pull request from an enthusiastic junior: review it, test it, own it.

Fundamentals matter more, not less

When generating code is cheap, judgement becomes the scarce skill. Knowing how HTTP, databases, rendering and concurrency actually work is what lets you tell a good suggestion from a plausible-looking one.

A simple workflow that works

1. Write the interface and a couple of tests yourself.
2. Let the assistant fill in the implementation.
3. Read every line. Run the tests. Refactor for clarity.

The tools will keep improving. The habit of understanding what you ship should not change at all.