Meta Coding Interview: What Gets Asked

The Meta coding interview is two problems in 45 minutes per round in CoderPad, both usually medium, with code execution turned off. A new grad or mid-level candidate gets one coding round. A senior gets two of four. Meta reuses a curated internal problem set, so its questions repeat far more than at other companies, and drilling the most-frequent ones is the highest-yield prep you can do. Since late 2025, one round is often an AI-assisted coding interview, covered below.

Coding is the round that decides most Meta loops. A new-grad (E3) or mid-level (E4) candidate runs a coding-dominated loop, often just coding plus a behavioral round, so the coding carries almost the whole interview. A senior (E5) candidate gets two coding rounds out of four. Either way, you cannot quietly underperform on code and make it up elsewhere.

Meta reuses its problems. The pool is curated and internal, interviewers pull from it again and again, so the same questions surface far more often than at other companies. By candidate reports, Meta's top problem shows up about three times as often as Amazon's most-repeated one. A generic algorithm syllabus wastes that edge. These lessons are built around the questions Meta actually keeps asking.

How a coding round runs

A Meta coding round is 45 minutes in CoderPad, a shared online editor. Two problems, both usually medium-tier. After a few minutes of intros and setup, you have around 35 minutes of real coding for the pair, so roughly 17 minutes each. The clock is the hard part. One elegant solution doesn't win you the round. You need two working ones, fast.

Because you check the code by eye, the bar on bugs is forgiving in one direction and brutal in the other. A small slip you catch and fix while tracing a sample input, an off-by-one or a missed null check, rarely sinks you. An algorithm that never reaches a working answer almost always does. Reach a complete, optimal solution and clean it up later. A beautiful half-solution that runs out of time loses.

The Meta angle

Finish both problems. The interviewer is calibrating your speed against a pool of candidates who clear both, so two complete solves is what you aim at. Not an absolute gate, though. People have passed with one finished solve and a credible start on the second. What actually gets scored is a complete, working solution explained clearly and fast, not a flawless one. So set the bar high in practice. Can you solve a known medium in about 15 minutes with a clean approach? Can you name the optimal one even when you don't finish coding it, and keep talking the whole way? Pick a problem off the frequency list, set a 17-minute timer, and write it in a plain editor: no autocomplete, no run button.

What earns a Hire

What Meta discourages: Dynamic Programming

Meta internally discourages dynamic programming, which is exactly where the "Meta doesn't ask DP" reputation comes from. That makes it lower-yield prep than the families below. But the norm is enforced unevenly across interviewers, so DP still shows up in a real slice of loops. It is usually a lighter single-dimension problem like House Robber, Coin Change, or Word Break. Sometimes it is Trapping Rain Water or the longest increasing path through a grid. What you will rarely meet is a gauntlet of hard, multi-state recurrences, since interviewers favor problems they can grade inside a tight 17-minute window. There is no full DP lesson in this section because the return on it is low, not because the topic is off-limits.

The Meta angle

Deprioritize DP for Meta, but do not go in blind to it. Hours spent grinding knapsack variants and edit-distance recurrences are hours not spent on string parsing and tree traversal, which is where Meta concentrates its questions. Keep a few of the common single-dimension patterns warm, like House Robber, Coin Change, and Word Break. That way a stray DP question does not catch you cold. Then pour the rest of your time into the families below. If you are also interviewing where DP is heavily tested, keep that on a separate track.

The Meta AI coding interview (the AI-enabled round)

The Meta AI coding interview, which Meta calls the AI-enabled round, puts an AI chat assistant open beside your editor. You are expected to use it. Meta began rolling it out around October 2025, and by late 2025 it appears in most loops, usually replacing one of your two coding rounds. What gets graded is not whether you can prompt a model. It is your judgment about when to trust the tool and when to overrule it.

It runs about 60 minutes in CoderPad, with the AI chat panel open alongside your code. You pick from a small menu of models, but the assistant is deliberately limited. Candidates consistently describe an in-house tool that is weaker than the ChatGPT or Claude they use at their desk. It's a capable helper that reads the files in your workspace and answers questions, but it cannot edit your code for you. Unlike the classic round, this CoderPad runs your code, with buttons to run the tests. The arc is usually fix a failing test, build something of around 120 or more lines, then optimize it.

What does the Meta AI coding interview test?

Your judgment about the tool, not your prompting. The assistant is deliberately limited: candidates consistently report the in-interview model is weaker than the ChatGPT or Claude they use at their desk. It will not name the bug for you or hand back a finished solution, because Meta tunes it not to. Lean on it for quick lookups and scaffolding. Solve the core yourself. Trust it where it is reliable, like recalling a library signature or scaffolding boilerplate. Stay skeptical where it is not. That covers a subtle bug that only shows up when two tasks run at once, or a confident complexity claim. You still own the design and the final code, and you are expected to understand and defend every line, including the ones the model wrote. Ask the model to explain a suggestion and check that explanation against what you already know. It reads far better than copying its output on faith and finding the bug after you have committed to it.

The Meta angle

The skill here is knowing when to overrule the AI. Meta wants engineers who ship with these tools every day, so what gets tested is judgment, not whether you can prompt. The trap is over-trusting it. Accept a plausible answer that quietly fails on an edge case and you score below the candidate who caught it. Practice with the assistant on, building a small feature, and audit every suggestion before you keep it.

What problems does the AI-enabled round use?

Not LeetCode questions. From recent reports, the pool is small and project-shaped. You get a multi-file codebase like main.py, test.py, solve.py, utils.py plus a data/ folder. You fix the failing tests, then fill in a stubbed function. The most consistently reported task is a partially built game whose missing logic and final strategy you complete. Reports differ on the exact problems beyond that. Some describe picking a subset of words whose letters do not overlap. That turns NP-complete once you optimize it. Others describe a card game where you repeatedly pick three cards that sum to fifteen. Treat those as illustrative, not a fixed list. The shape matters more than the specific problem: read an unfamiliar codebase fast, get the tests green, then reason honestly about what your solution costs.

How to prepare for the Meta AI coding interview

Practice with an assistant on, not off. Open a multi-file project you have never seen, get its failing tests green, then add a feature of a hundred lines or so while a model answers your questions beside you. Then deliberately audit it: paste back a function with a planted off-by-one or a race between two tasks, read it critically, and decide fast whether to keep it or throw it out. Reading suspect code quickly under pressure is the muscle this round measures. Rehearse the classic format alongside it. Most loops still pair the AI round with a standard two-problem coding round, and both reward the same core skill: reaching a correct, optimal solution fast. So the frequency lists in this section stay your highest-yield prep either way.

The patterns that show up

Across roughly a thousand recent Meta interviews, the spread leans hard toward one family. String and stack parsing sits at the top, the problems where you scan characters and track state on a stack. They're Meta's signature. The chart below buckets the most-asked problems into pattern families so you can see where to start.

Strings and stacks lead, with trees close behind and the back half clustered. This section walks the families in that order. Each lesson opens with a short read on how Meta frames the pattern and the variant they prefer. Then it hands you the most-asked problems as runnable exercises with worked solutions. The deep algorithm walkthrough for each lives in the free Foundations course, linked from every lesson, so these pages can stay on the Meta-specific layer.

If a pattern's underlying algorithm is new to you, start with the matching lesson in the free Foundations DS&A course and come back here for the Meta framing. The Foundations course teaches the technique from scratch; these lessons show you the exact way Meta asks it.