Studying at the Faculty of Computational Mathematics and Cybernetics at Moscow State University gave me not a set of formulas but a habit: getting a solution isn’t enough — you have to defend it. What are the assumptions? Where will it stop working? Why was this particular algorithm chosen?
With AI this habit is especially useful. It easily writes convincing code that hides a wrong model of the problem. The code looks neat, a test on one example passes — and in production it breaks on empty input, duplicates or concurrent requests.
After generating code, I don’t ask “check the code”. I ask for a mini defense:
You are a strict examiner. Ask me 7 questions about this solution: about assumptions, edge cases, complexity, data errors and concurrency. After each question, write what answer would be sufficient.
Then I answer myself or ask AI to answer with references to specific lines. If the answer is vague, that’s a place to improve.
Why it works: AI switches from “create quickly” mode to “find grounds for the solution” mode. And you check understanding, not syntax.
You don’t need this for a twenty-line one-off script. But for logic that will be maintained, it almost always pays off.