Practise on realistic, unfamiliar code under a timer, with the kind of tool you will be allowed, and review how you worked afterwards. Interviewers who allow AI still grade problem solving, verification and communication, so rehearse planning before prompting, giving the AI the facts it cannot guess, running tests after each change, and saying why you chose one design over another.
What should you practise on?
Code you did not write. Meta's AI enabled round, per Hello Interview, starts with a bug in an existing multi file project. Google's pilot, as reported, is about reading, debugging and optimising existing code. Canva uses realistic, deliberately ambiguous tasks such as an airport takeoff and landing controller. Short algorithm puzzles do not train these skills.
Our three practice rounds are built this way: a small service to extend with rate limiting, an invoice module to debug, and a notes app that needs sharing. Each has 6 to 12 files of existing code, tests that fail until the work is done, and a production trap the brief does not mention.
How should you run a practice round?
Set a 60 minute timer before you open your AI tool and stop when it rings. Use your own agent, the way you would at work. Read the brief and the code first, write down the decisions the brief leaves open, and keep a short notes file of what you decided and why.
When the timer stops, run the round's extra checks. They test what a reviewer would test, including the trap. Then review the session, which is where most of the learning happens.
Which habits matter most in the room?
Break the problem down and hand the AI one step at a time. In a study of 19 Microsoft developers, issues solved incrementally succeeded 83% of the time against 38% when the whole issue was handed over at once. Give the AI the facts it cannot guess: in the same study, issues where developers supplied expert knowledge succeeded 64% of the time against 29%.
Verify as you go: run the tests after each change and read the diff before you accept it. Correct precisely when the AI goes wrong, and make small fixes yourself; Shopify's head of engineering calls prompting for a one character fix a warning sign, and in the Microsoft study writing some code by hand went with 73% success against 36%. Finally, say why you chose what you chose.
Should you practise with the same tool as the interview?
Practise with the same kind of tool. Meta's assistant reportedly answers in a chat and cannot edit files, so rehearse with a chat only assistant and make every edit yourself. Google's pilot uses Gemini. Canva lets candidates bring tools such as Copilot, Cursor or Claude. If you only ever practise with an agent that edits and runs everything, a chat only round will feel slow.
Ask your recruiter which tool, which interface and how long you will have. Then practise in that setup at least twice.
How do you review a practice round?
Each round includes a REVIEW.md with the five part practice rubric and a prompt to paste into your own agent. It asks the agent to read your session transcript and quote the moments that helped or hurt on each dimension, and to say when there is not enough evidence. The transcript stats tool adds plain counts, such as how many prompts you wrote and how often you ran the tests.
Look for moments, not a number: the prompt where you handed over too much, the edit you accepted without reading, the check you skipped. Then do another round and watch for the same moment.
What should you avoid?
Pasting the whole brief into one prompt. Accepting a change that edits a test to make it pass. Asking the AI to fix something you could fix faster by hand. Going quiet while the AI works, when the interviewer needs to hear your reasoning. And skipping the unaided rounds: most loops that include an AI round also include classic ones.