We use analytics and advertising cookies to understand how the site is used and whether our ads on Facebook, Instagram and Microsoft search (Bing) work. They are set only if you accept. See our Privacy Policy for details.
Home / Meta AI-enabled coding interview
Meta now hands you an AI assistant in one of its onsite coding rounds and watches what you do with it. The candidates who struggle are rarely the ones who cannot code - they are the ones who accept what the assistant writes. This page covers the format, what gets graded, and a drill you can do right now.
Do you understand and break down the problem before handing pieces to the assistant?
Are your asks specific and incremental, with the context the model needs?
Do you read, run and question its output - and catch it when it is wrong?
Does the final code reflect your judgment, or is it pasted wholesale?
The ask: “Add GET /users/search?q=<name>&page=<n> that returns 20 users per page whose name contains q.”
Context: Flask app on SQLite via the standard sqlite3 module. The endpoint is public.
Click any added line you would block in review and say why. This change has 2 planted defects.
| 41 | @bp.get("/users/<int:user_id>") | |
| 42 | def get_user(user_id): | |
| 43 | return jsonify(fetch_user(user_id)) | |
| 44 | ||
| 45 | + | |
| 46 | + | |
| 47 | + | |
| 48 | + | |
| 49 | + | |
| 50 | + | |
| 51 | + | |
| 52 | + |
Nine more drills in Python, TypeScript, Go, Java and SQL: all PR-review drills. Then rehearse the whole round with AI-assisted coding practice - a real assistant, a real code runner, and a scorecard on the four dimensions above.
Since October 2025 Meta's E5-E7 software engineering loops have replaced one onsite coding round with a round where you are given an AI assistant and expected to use it. Candidates describe roughly 60 minutes on one extended problem in a multi-file CoderPad project, with a choice of models in a chat panel. You are graded on how you work with the assistant, not only on the final code.
How you frame and break down the problem before delegating, how specifically you direct the assistant, whether you verify what it produces - reading it, running tests, catching its mistakes - and whether the final code reflects your own judgment. Pasting generated code without checking it is the failure the format exists to catch.
Rehearse the workflow, not just algorithms: plan before you prompt, ask for small pieces, run and read everything that comes back, and practice reviewing AI-written code until spotting a missing error check or an injection is reflex. gitGood has both halves: AI-assisted coding sessions with a scorecard, and code review drills built from AI-written diffs with planted bugs.
Yes. The AI-enabled round replaces one coding round, not all of them, and other rounds and other companies still test unaided problem solving. You also cannot verify code you could not have written - the assistant makes you faster only if you can tell when it is wrong.