It is a 60 minute onsite round in a three panel CoderPad: project files on the left, the editor in the middle and an AI chat on the right. You get one multi file problem in three phases: fix a bug, build the core feature, then make it hold up on bigger inputs. According to Hello Interview, the AI can read the project but cannot edit it.
What happens in each phase?
The first phase is bug fixing. Hello Interview reports that the rules on AI in this phase vary by interviewer, and that it can include explaining the existing tests and the complexity of the existing code. Expect to read before you write.
The second phase is the core implementation, and here AI use is explicitly allowed and expected. The third phase is optimisation: larger test inputs expose the weaknesses of what you built in phase two, so a solution that only just passed the small tests will be found out.
What is the assistant allowed to do?
It has the project as context and answers only in the chat. It cannot change files, so you write or paste every change yourself. The guide lists models from several providers to pick from and reports that the list can differ between candidates. Supported languages, per the same guide, include Java, C++, C#, Python, Kotlin and TypeScript.
Because the AI cannot run anything, verification is your job. You decide which tests to run, read the output, and tell the interviewer what you checked.
What are interviewers looking for?
The reported scoring uses the same four competencies as Meta's classic coding round: problem solving, code quality, verification and communication. The guide is clear that the round is not about prompt wording. What earns credit is understanding the code, checking what the AI proposes, and explaining why you kept or rejected it.
The negative signals reported by candidates are the mirror image: leaning on the assistant until the quality of the solution slipped, or being unable to explain code that came from the chat.
How is it different from the classic coding round?
The classic round is a short, self contained algorithm problem with no assistant. The AI enabled round gives you a small project with several files, existing tests and existing bugs, which is closer to a first week on a real team. Reading unfamiliar code quickly matters more, and remembering a textbook algorithm matters less.
Both rounds stay in the loop, so you still need to practise the classic format. Preparing for one does not replace the other.
What do candidates report going wrong?
Two things come up in the guide. Some candidates found the assistant less helpful in the real interview than in practice sessions; the guide's leading theory is a change in the assistant's instructions, which is unconfirmed. And some candidates spent too long negotiating with the chat instead of making a small change themselves.
Both point to the same habit: decide what you want before you ask, keep requests small, and type the obvious fix yourself.
How can you rehearse it?
Set a 60 minute timer, open an unfamiliar codebase, and use your assistant in chat only mode, so that every edit goes through your hands. Our invoice totals round is a debugging task on existing code with a production trap the brief does not mention, and the rate limit round is an extension task on a small service.
When the timer stops, run the checks, then review the session with the practice rubric. Look for the moment you stopped reading and started accepting; that is usually where the points went.