Meta coding interview questions to expect

You are looking for the real question list so you can stop guessing what to grind. This section names the families Meta actually reuses, and it is honest about what a list cannot do. Patterns transfer; a screenshot of someone else's prompt usually does not.

• Arrays and strings. Two pointers, sliding windows, prefix sums, and in-place rewrites.
• Graphs. BFS and DFS on grids or adjacency lists, connected components, and shortest paths on unweighted graphs.
• Trees. Traversals, lowest common ancestor, serialization, and path aggregates.
• Heaps and ordering. K-th elements, merge K lists, and streaming top-K.
• Intervals and maps. Merge, overlap, and sweep-line shapes that look simple until the edge cases.

Posted items go stale because interviewers pick their own prompts. What stays stable is the family. If you can name the pattern, state the complexity, and dry-run a failing case out loud, you are closer to the round than someone who memorized last month's screenshot. Role-specific banks sit under the question banks topic.

The pad is the same object in both stages: a shared editor, cameras on, often no execution. The difference is density and follow-up. A phone screen is usually one interviewer and two problems with a little more room to recover. A loop coding round is the same shape under a tighter clock, with a second interviewer sometimes sitting in, and with less patience for a silent start.

Both stages want a working approach before code appears. Clarify the input, pick a structure, estimate time and space, then write. The second problem is often a twist on a familiar family rather than a new domain, which is why finishing the first problem with time left matters more than polishing it.

How this round sits in Meta's full pipeline is on the company interviews topic.

Practice the round, not only the puzzle. Solve in a plain editor with highlighting off, speak every choice, and cap the session at 45 minutes with two problems. Hand-trace your code, because the pad will not run it for you. Record one session and listen for the silences; those are what the interviewer hears.

A backend engineer targeting an E5 role spent three weeks on a public Meta-tagged list, then froze on a graph follow-up that was not on it. She recovered by restating the new constraint, picking BFS, and talking through the visited set. The interviewer scored that she could move, not that she had seen that prompt.

Timed, narrated reps against an AI interviewer are on the mock interview page. That rehearsal is for the spoken part of the pad, which is the part silent grinding never trains.

The coding pad is a shared surface. Keystrokes, deletions, and pauses are visible to the interviewer, and a screen-shared IDE puts anything on your display inside the share. Live assist is out of scope on that surface, including proctored or company-managed devices.

SubcueAI is built for the spoken rounds around that pad. The native macOS and Windows app captures system audio plus your microphone into a floating local overlay. A browser extension Side Panel does live assist for a meeting tab on Chromium browsers, capturing only that tab's audio, the interviewer, never your mic. No meeting bot joins the call, and nothing is injected into the meeting page.

Use it to keep structure on a behavioral or design conversation. Do not use it on the pad. Limits and architecture are on the detectability topic.