What candidates say the round looked like, summarised in our words, each linked to where it came from. LinkedIn hasn’t confirmed these details unless the source says “Official page”.
Background on the format: our LinkedIn guide.
Had this round? Add your report →
Done the week before: the coding with AI round was the most unpredictable of the loop. It was an algorithmic problem of the LRU cache kind with 3 or 4 extensions. You must come up with the core logic yourself; asking the AI for the algorithm doesn’t work. Use the AI to generate test cases and edge cases and to stress test. Mostly they want to see that you stay in control of the generated code: build it step by step and delegate small parts in iterations.
Interviewer’s view: in LinkedIn’s AI coding round you get both Sonnet and Opus, but Opus is usually too slow to finish in time. A one-prompt solution can be fine and still lead to a hire, but questions are deliberately a little vague, so prompts need a lot of detail, either up front or through follow-ups. Either way, interviewers check whether you reviewed the code by hand and whether the generated code is production ready.
Coding with AI in CoderPad on a scheduling problem built around an interval tree. The tooling went badly: the AI panel hung, it took over 30 minutes to get the pad into a usable state, one answer froze for 4+ minutes, and the question and starter code were out of sync until a full reset. The interviewers were patient, and the candidate still wrote the answer with the AI and ran it.
Senior engineer loop: a two-pointer coding round with two harder follow-ups, then the AI coding round, where the task was to go through the files and implement small functions. The candidate’s verdict on the built-in chatbot: it was bad, so don’t waste much time on it. The loop also had an API quota system design round and a manager round.
Staff SWE onsite, coding with AI round: a modified interval-merging problem in 3 incremental parts. The interviewer watched how the prompts were written and whether the candidate verified that the assistant’s solution was correct, and an optimal solution carried a lot of weight. The candidate was rejected.
AI Engineer virtual onsite, first round: an hour in an AI-assisted IDE. The task was to implement an LRU cache and then add key validation. The candidate had the AI generate the skeleton, a hash map plus doubly linked list, and prompted Claude for a custom linked list class. The interviewer kept asking for more features, and the solution got over-engineered through prompting alone. The candidate wasn’t sure what the round was grading, and the loop ended in a rejection without feedback.