How to Use AI as a Software Engineer Without Breaking Things

You've probably tried a few AI coding tools by now, like Cursor or Claude Code. A lot of the advice on using them, though, comes from the companies that make them or from people online who promise they'll make you 10 times faster.

For this piece, I asked three engineers how they use AI at work: Pooja Naik, Senior Software Engineer; Viraj Patel, a software engineer who joined Microsoft about a year ago; and Tom Archer, Independent AI/ML Research Engineer at Signal & Syntax, who has been programming professionally for more than 40 years.

They're all at very different points in their careers, but they follow a similar rule: AI helps with the work, and they make the important decisions themselves. It also means more of the job is now checking code. As Archer put it, engineers are "increasingly reviewing and taking responsibility for code they didn't personally write."

Here are six habits for using AI as a software engineer, based on what they told me.

GTM Engineer Salary and How to Become One in 2026

We spoke to Kaan Un, who recounts how AI is helping him do the work of an entire team and why he thinks the role could become a six-figure career.

TechloyDavid Adubiina

/1. Don't take confident answers at face value

Naik was working through an unfamiliar part of her codebase when she asked an AI tool where a particular function was used. It told her, confidently, that it wasn't used anywhere and was dead code she could remove.

Naik knew from experience that the function was still being called, so she pushed back. But the AI stuck to its answer. She pushed again, and this time it checked more carefully and admitted it had missed where the function was used.

"If I hadn't already known that code path, I would have deleted something that was actually load-bearing," she said.

Something similar happened to Patel. AI once gave him a fix that looked right but ignored an important edge case. He caught it by testing the fix and comparing it against the actual business requirements. He treats AI as "a helpful assistant, not an authority."

Archer says he has caught AI being wrong "countless times." The problem, he said, is that today's models are "extraordinarily good at sounding authoritative," and "confidence and correctness aren't the same thing."

So before you act on an answer, test it or check it against something you know, the way Naik and Patel did. And don't assume one round of pushback is enough.

/2. Use it to think through problems with you

Try asking AI to point out the weak spots in your plan before you ask it to write anything. That's how Archer works. He no longer thinks of AI mainly as a code generator, and treats it more like a colleague whose role changes depending on what he needs.

"Sometimes it's a pair programmer. Sometimes it's a research assistant. Sometimes it's effectively a rubber duck that talks back," he said.

He'll sometimes ask it to find problems with an idea he's had and tell him what he's missing.

Naik starts the same way. She uses Claude or ChatGPT to think through her approach before she writes code, not to write it.

Patel saw the difference this makes when he was a teaching assistant at Syracuse University. The best students used AI like a tutor to understand concepts and debug problems. The ones who struggled sometimes used it just to get answers. That worked in the short term, but later they found it harder to explain their own solutions.

/3. Use it to learn code you didn't write

When Naik had to work on a part of her codebase she'd never seen before, she asked an AI agent to trace through the code and explain how the system worked.