ChannelPulse
The AI-assisted coding round, phase by phase
The structured AI round is sixty minutes in a multi-file repo with an AI chat beside the editor, run as three phases on one problem. This lesson walks the format minute by minute, says what each phase is grading, and shows how much of it you actually need to finish.
What you'll be able to do
- Describe the structured AI round's layout and the three phases on one problem
- Budget sixty minutes across orientation, bug fix, implementation and optimisation
- Say what each phase grades, and why fixing the first bug yourself is a signal
- Recognise the problem families these rounds draw on, so you can drill the patterns underneath
Before this: what-changed
Meta's version of this round is the best documented, and the other structured formats (LinkedIn's, Amazon's assessment, Uber's) share most of its shape, so it's the reference used here. Treat the specifics as current as of late 2026; companies re-test their questions with every major model release and adjust.
The layout
Three panes: a file tree, an editor, and an AI chat. The assistant can read every file in the repo but cannot edit any of them. Whatever it writes, you paste in yourself, which is a deliberate design: every line that ends up in the code passed through your hands.
The model menu has changed over time, and Meta's FAQ now lists models from several labs. Several candidates report that the assistant was noticeably less helpful in the real interview than in the practice environment, describing code instead of pointing at bugs. That's reported, not confirmed; plan as if the tool is mediocre and you'll never be caught out.
A former Meta staff engineer works a Meta-style problem in one take. The chapters line up with this lesson's phases; the code-comprehension stretch is the part most candidates skip.
Jump to the part you need
The sixty minutes
| Minutes | What happens | What it grades |
|---|---|---|
| 0–6 | Orientation: read the files, the tests, the data | Whether you can find your way in code you didn't write |
| 6–20 | Phase 1, bug fix. One or more small, non-algorithmic bugs: a wrong comparison, an off-by-one, a type mix-up, a safety cap where a visited set was needed | Independence and reading. Some interviewers ban the AI here; you may be asked to explain the tests and the complexity |
| 20–45 | Phase 2, implementation. The core feature. Reported at 120+ lines, harder than a typical medium problem. AI use is expected | Control: your approach, small delegated pieces, review of each |
| 45–60 | Phase 3, optimisation. Larger input files that break the phase-2 solution, sometimes needing a different algorithm | Measurement and judgment |
Two reported facts change how you spend the hour. First, candidates who passed describe hitting three or four checkpoints and not finishing phase 3; how you work is graded throughout, so a calm, verified phase 2 beats a rushed, broken phase 3. Second, the most common regret is skipping orientation. Five minutes of reading makes every later prompt shorter and every review faster.
Phase 1: why fixing it yourself is a signal
The first bug is chosen to be findable by reading. Asking the AI to find it is allowed in many formats, but it costs you the clearest early evidence that you understand the code, and it reads as dependence. Fix it by hand, run the whole suite (not just the failing test), and say what the wrong fix would have been — raising the cap, wrapping the error in a try/except.
Phase 2: you decide, it types
This is the phase the round is named for. The pattern that passes, across every source: name the approach and the algorithm yourself, then hand the assistant small, specific pieces you can review in one read, and run the tests after each. The next lesson is about doing that well; the short version is that the interviewer should never see the AI make a decision.
Phase 3: measure, then change
The large inputs exist to break a correct-but-slow phase 2. Time the parts before guessing, find the real cost (a list used as a queue, copied paths, a set of tuples where an array would do), and explain it before you change it. If the fix is a different algorithm (greedy to trie, depth-first search to a bitmask), say so plainly — that's the judgment being tested.
What these problems are made of
The problems are picked so the assistant can't one-shot them, and they cluster in a few families: graph search and pathfinding, topological sort, backtracking, greedy and bin-packing, dynamic programming, parsing, and data-structure design. Reported Meta problems include a maze solver, choosing words to maximise unique characters, a card game, and friend recommendations.
| If the problem is… | The pattern you should name | The trap |
|---|---|---|
| Shortest route through a grid | Breadth-first search | Keeping the depth-first search the repo already has |
| Order tasks with dependencies | Topological sort, cycle detection | Running tasks in file order |
| Pick a subset under a constraint | Backtracking with pruning, or greedy with a proof | Greedy that fails on one adversarial input |
| Recommend by shared connections | Counting over neighbours, a heap for top-k | Sorting everything for k = 5 |
Before the day
Do the company's practice environment if there is one; Meta's is reportedly harder than the real questions. Practise the whole arc against a sixty-minute clock in a repo where the AI can read but not edit. And practise at least once with a weak model or no AI at all, so a disappointing assistant in the real round is an inconvenience rather than a crisis.