Meta's AI-Assisted Coding Interview: Format and Prep

Since late 2025, candidates in Meta's onsite loop have been reporting a coding round that looks nothing like two LeetCode problems in 45 minutes. In the Meta AI-assisted coding interview you get an hour, a small project you did not write, a test suite and a chat panel with an AI model in it. If you prepared only for the old format, the file tree alone can cost you the first ten minutes.
A warning before the detail. Meta's public prep page for the full loop, as of November 2026, describes the onsite only in general terms and points to a PDF guide; it does not spell this round out. Everything below comes from prep sites and collected candidate accounts published between April and November 2026. They agree on the outline and disagree on several specifics, which are marked below. Meta can change the format at any time, so confirm with your recruiter what your loop contains.
What changed and when
Accounts date the first interviews to October 2025; one site that collects candidate reports has its earliest one from 17 October 2025. The new round replaces one of the two onsite coding interviews. The other stays as it was: no AI, classic problems.
Beyond that, sources split.
- Pilot or standard. One guide still calls it a pilot with candidates picked at random and a full rollout expected during 2026. Another, from April 2026, says flatly that it is no longer a pilot. A third says it is rolling out for back-end and operations-focused roles.
- Who gets it. Reports cover software engineers, machine learning engineers and engineering managers, from new graduates up to E6. Two guides say candidates at E7 and above have a single coding round and it is this one.
- Where in the process. Almost all reports are from the onsite. One collection includes two from a technical screen.
Recruiters are reported to say in advance which coding round is which. If yours does not, ask. Meta is not alone in this: the guide to AI in coding interviews compares what other companies allow and forbid.
The environment
The round runs in CoderPad, set up to feel closer to an editor than to a blank pad. The description that recurs is three panels: a file explorer on the left, the code editor in the middle, and the AI chat with the problem instructions on the right. There are tests you can run whenever you like. One guide mentions a terminal for output; another says code reruns on save and you pick tests or a main program from a menu.
You can choose the model and switch during the round. The lists differ by source and date, which suggests the menu changes: guides name versions of GPT, Claude, Gemini and Meta's own Llama, and one says Llama is the default. Two candidate reports describe a model that stopped responding mid-round. Do not build your plan around one model's habits.
The detail with the most effect on how you work: two guides say the assistant can only answer in the chat panel and cannot edit your files, so every line still passes through your hands. A third notes that CoderPad itself offers both a read-only mode and an edit mode and that it is not clear which Meta turns on. Plan for chat only and treat anything more as a bonus.
Languages reported include Python, Java, C++, C#, TypeScript and, in one list, Kotlin.
The task shape
You get one problem for the whole hour, in parts that build on each other. The codebase is described as a few hundred to a few thousand lines across several files, with existing classes, data models and tests. The interviewer spends the first five minutes or so on orientation.
Most accounts describe three phases.
- Read and repair. You start by finding and fixing a bug in the existing code, sometimes also explaining what the tests check or the complexity of a function. Rules on AI here vary by interviewer. In one collection of reports, three candidates had it switched off for this part.
- Build the main piece. A substantial implementation on top of the existing code. Using the assistant is allowed and, by most accounts, expected. One guide estimates more than 120 lines and calls it harder than a medium LeetCode problem.
- Make it hold up. New rules, bigger inputs or tiered tests that expose weaknesses in what you just built.
Not every report fits three phases. A site that counted parts in one recurring problem found four parts in eight reports and five in one. Another guide frames the round as one of three scenarios: add a feature, finish a half-built one, or debug a failing one.
Problems candidates name include a maze solver that has to print its path, a card game that looks for three cards summing to 15, picking words to maximize unique characters, and a friend recommendation feature. One guide passes on an insider's estimate that the pool is small, around nine problems. Chasing leaked problems is a poor use of time. The interviewer watches how you read unfamiliar code, and recognition does not help with that.
What is graded
Two sources list four things, and one of them says they are the same four used in Meta's traditional coding round:
- Problem solving. Do you understand the problem and reach a sound approach.
- Code quality. Is the code clean, and do you understand all of it, including what the assistant wrote.
- Verification. Do you run things, test edge cases and check the assistant's output before trusting it.
- Communication. Can the interviewer follow your reasoning.
One guide gives three areas instead and folds communication into the rest: problem solving, code development and understanding, verification and debugging. The same guide reports Meta's position as being that AI use is optional and does not by itself affect your score. That fits the rest of the picture. You are graded on engineering, and the assistant is a tool you may use well or badly. It is also practical advice in disguise: by most accounts the hour is tight without help on the mechanical parts.
The round is reported to include questions the assistant cannot answer for you in any useful way: why you would change a function signature, what the runtime of your approach is, which data structure and why. The interviewer sees your whole chat. If the model chose the hash map, you still have to defend it.
Working with the assistant
The candidates who do well seem to treat the model like a fast junior colleague: good at typing, unreliable at judgment.
Ask it for these.
- A summary of a file or class you have not read yet, which you then check against the code.
- Boilerplate: parsing, a helper, a test scaffold, a loop you could write in your sleep.
- A second opinion on an error message.
- Edge cases you may have missed, as a list to consider.
Keep these for yourself.
- The approach. Decide the algorithm and say it to the interviewer before you prompt.
- The reading in phase one. Skipping it is the most commonly reported mistake, and it costs you later when you cannot place new code.
- The final check. Read every line you paste. Run the tests after every change.
Ignore it when it confidently proposes something that does not match the existing code's conventions, when it starts rewriting code that already works, or when you have gone two rounds of prompting without progress. At that point write it yourself.
A prompt that hands over the whole problem statement is the pattern guides warn about most; one says large unreviewed pastes are read as a red flag. Small, scoped prompts are better. An invented example of the difference, for a maze task:
- Weak: "Here is the problem. Write the solution."
- Better: "In
Maze.neighbors, gates should only be passable in the direction stored inGate.direction. Write that check. Do not change the method signature."
The second one shows the interviewer that you read the code, know where the change goes and set a constraint. Then you read what comes back, say aloud what it does, and run the tests.
Mind the opposite failure too. One guide lists spending too long reviewing AI output among the mistakes. Review in proportion: a five-line helper gets a glance and a test, a 60-line core function gets a proper read.
A practice routine
Solving algorithm problems in a single file does not train this. What you need is repetitions of reading someone else's code under time pressure with a chat assistant beside you.
- Get the real environment if you can. More than one source says recruiters can provide a practice CoderPad session with the AI panel enabled. Ask. One guide notes that the practice puzzle is harder than most real problems, so do not panic at it.
- Build three or four small projects to practise on. Take an open-source repository of a few hundred lines, or have a model generate a small multi-file project with a planted bug and failing tests. Do not read it beforehand.
- Run the hour. Five minutes reading with no AI. Fix the bug. Then implement a feature with a chat-only assistant, typing or pasting every change yourself. Then add a harder constraint. Stop at 60 minutes.
- Talk the whole time. Record it. The habit to break is going quiet while you read the model's answer.
- Do one session with the assistant off. A model may stall on the day, and some interviewers switch it off for part one.
- Afterward, write down every line you pasted without understanding. That count should reach zero.
What you cannot produce alone is the interviewer: someone who stops you to ask why, and whom you must keep talking to while a chat window competes for your attention. Do a couple of sessions with a colleague watching. If you would like that person to have run coding interviews, book a mock with a coding interview coach and ask them to set it up this way, with a small codebase and an assistant open.
What still needs classic prep
For E6 and below, the other coding round is unchanged by all accounts: two problems, no AI, speed and correctness. Nearly every report of the AI format comes from the onsite, so expect the technical screen to be classic too. One site suggests about twenty minutes per problem as the pace to train for.
So this format adds to your preparation and replaces none of it. Data structures and algorithms still decide half of your coding signal, and they decide whether you can judge what the assistant hands you in the other half. Someone who cannot write a breadth-first search unaided cannot tell whether the generated one is right.
The rest of the loop is unchanged as well: a design round, covered in the Meta system design guide, and a behavioral round, covered in Meta behavioral interview questions.
Mistakes candidates report
- Starting to code or prompt before reading the existing files.
- Pasting the full problem into the chat and copying the answer.
- Pasting code they cannot explain when asked.
- Skipping the tests, or running the whole large suite early and drowning in output.
- Talking to the assistant and forgetting the interviewer.
- Spending so long on phase one that phases two and three never start.
- Trusting a confident wrong answer because checking felt slow.
Most of these come down to pace and attention, and both are trainable in a week of honest practice hours.
Where coaching fits
A coach is useful here for the part you cannot simulate: being watched and questioned while you work with an assistant. Nine of the 40 coaches on techinterview.coach list coding interviews as of November 2026, and you can see them on the coding interview coaches page; say in your booking note that you want the AI-assisted format so they can prepare a codebase. Three coaches have Meta on their work history and appear on the Meta page. For the talking half of the skill, the practice interview has a coding talk-through round with an AI interviewer and a written recap. It is free; sign in to start.
Related guides
- Meta system design vs product architecture: the design round in the same loop and which variant you will get.
- Meta behavioral interview questions: questions by signal for a round that weighs heavily on level.
- Meta behavioral interview: how that round runs and what the interviewer writes down.
- Mock interviews guide: how to set up practice that includes a second person.
- Google system design interview: includes where Google stands on AI tools in interviews, for comparison.