Screens, OAs & Phone Rounds

The recruiter screen: 30 minutes that set your level

The recruiter screen is a 20–30 minute call covering your background, motivation, logistics (location, visa, timeline), and compensation expectations. It is non-technical — but it is not low-stakes, because the recruiter calibrates your target level from this call. In Google's 2026 pilot, some recruiter screens even include a short technical discussion of a past project's design.

Prepare three things:

1. A 90-second narrative. Practice a tight version of your background that leads with ownership: what you built end-to-end, what you drove, what shipped. "I've spent three years at Acme, where I owned the payments-retry service — designed it, shipped it, and ran it in production" is a mid-level opener. A chronological list of jobs is not.

2. A compensation deflection. Know your market range from levels.fyi (a site that crowdsources real compensation data by company and level), but do not name a number. A working script: "I'd like to see the full offer first — I'm targeting market rate for L4." Naming a number early can only anchor you low.

3. Questions that change your prep. Ask: What is the exact loop composition for this role? Is AI tool use allowed in any round? What is the timeline? Companies differ enormously on all three, and the answers determine what you practice.

The online assessment: assume you are being watched

The OA is 1–2 timed algorithmic problems (or a work-sample-style task at Google) on a platform like CodeSignal or HackerRank, usually before you talk to any engineer. The critical 2025–2026 fact: these platforms are now heavily proctored, because cheating exploded. CodeSignal reported fraud attempts on proctored assessments more than doubling, from 16% in 2024 to 35% in 2025, and now layers digital, AI, and human proctoring. HackerRank shipped AI proctoring in July 2025.

What the systems actually detect:

Practical rules that follow directly from this:

• Do not use outside AI unless the assessment explicitly permits it. Some assessments now provide an in-platform AI assistant and expect you to use it — read the rules for each one. Outside that, detection is good enough that a flag ends your candidacy.
• Take it in a quiet room, webcam on, no second monitor. Remove the ambiguity before it exists.
• Type like a human solving a problem — because you are one. Comment your thinking, test, and debug normally. Ironically, honest problem-solving behavior is exactly what the proctoring models score as legitimate.
• Practice LeetCode-medium problems under a real timer, in the actual platform's environment beforehand. Platform editors have quirks; discover them before the clock is running.

The technical phone screen: talking is the skill

A 45–60 minute live coding session with an engineer over CoderPad or a video call: one or two algorithm problems, sometimes a practical debugging task. For mid-level, the bar is a working, reasonably optimal solution with clear communication — brute force alone does not clear it.

There is also a new dynamic to understand. In a survey of Big Tech interviewers by interviewing.io, 81% said they had suspected AI cheating in remote interviews, and 31% had caught it outright. Interviewers responded by probing harder: more "why did you choose that?" questions, more live modifications ("now make it handle streaming input"), camera-on norms, and — at several companies — final rounds moved back in person. The ability to extend and adapt your own solution is now the trust signal that separates genuine skill from a pattern-matched answer.

Drill this loop until it is automatic:

1. Restate the problem in your own words and confirm.
2. Enumerate edge cases out loud before coding (empty input, single element, duplicates, overflow).
3. State your approach and its complexity before writing code, then code it.
4. Trace your code line-by-line out loud with a real example when done.
5. Expect a follow-up variant — and treat it as the part being scored.

At this stage, one or two mock interviews (interviewing.io, Pramp, or a peer) buy you more than another fifty LeetCode problems. The gap for most mid-level candidates is narration under pressure, not algorithms.

The take-home variant: the review call is the filter

Take-homes are growing fast: roughly 68% of companies now use some take-home component (up 12% year over year), and about 41% run a hybrid model — a take-home followed by a live session where you defend and extend your submission. "Vibe coding" rounds — build a working prototype with Cursor, Claude, or Copilot, then walk through your decisions — spread from startups into larger companies through 2025–2026.

AI policy diverges sharply by company: Canva explicitly expects candidates to use AI tools in technical interviews, while Stripe prohibits AI in all interview stages including take-homes. So:

• Confirm the AI policy in writing before starting. One email: "Just to confirm — am I permitted to use AI coding tools for this exercise?"
• If AI is allowed, use it — but be able to explain and modify every line. The live review is where hybrid processes filter people. If you cannot extend your own submission fluently, the take-home reads as someone else's work.
• Ship something that runs, with a README covering trade-offs, what you would do with more time, and how you verified it. A polished core flow beats a wide spread of half-done features.
• Timebox to what they ask. Spending 20 hours on a "4-hour" exercise signals poor judgment, not dedication.