AI Coding
6h of practice10 problems
The AI-allowed round is now the most common new format and the least prepared for. It is not a test of whether you can get a model to write code; it is a test of whether you stay the author, whether you can say what you expected, notice when the output is subtly wrong, and defend any line as if you had typed it.
These problems are built so that the obvious generated answer is plausible and wrong in a specific way. Finding that is the exercise.
The round · Sixty minutes, your own tools open, thinking out loud. The interviewer rarely reads your code closely; they are listening to what you asked for, what you expected back, and how you found out it was wrong.
Difficulty
Topic
10 problems
Medium
6- Build a feature with a model, out loudThe baseline AI round. You are scored on what you asked for and what you expected back.Direction · Narration · Decomposition45 min
- Review generated code that is subtly wrongClean, confident, well-commented, and broken in three places.Code review · Verification · Concurrency30 min
- Have a model write the tests, then find what it missedGenerated tests cluster around the happy path and are judged by their count.Testing · Coverage · Edge cases35 min
- Find the security bug in generated codeIdiomatic, readable, and it will leak every other customer's data.Security · Code review · Authorisation30 min
- Debug a failing test with a model in the loopA flaky test, a confident wrong diagnosis, and the discipline not to take it.Debugging · Hypotheses · Context35 min
- Use a model for an estimate, and catch it being wrongConfident arithmetic, plausible units, and an answer three orders of magnitude out.Verification · Numbers · Judgement30 min
Hard
3- Refactor code you have never seen, with a modelAn 800-line file, no tests, and a model that will happily rewrite all of it.Refactoring · Safety nets · Comprehension45 min
- Plan a 400-file migration with a modelThe work is mechanical, the review is not, and 400 files cannot be read.Codemods · Scale · Review strategy45 min
- Review an LLM feature for prompt injectionThe model does exactly what it is told. The problem is who gets to tell it.Security · LLM features · Trust boundaries40 min