13 Interviewer Quirks — the weird things candidates don't see coming

Interviewers read a candidate's weakest round inside the first 90 seconds — and it's almost never the one they studied for.

LeetCode is a hammer. It's a great hammer. But 13 of the most sought-after tech companies have built interview rounds that are not nails, and candidates keep swinging anyway.

The StrongYes dossier set covers every major tech company's interview process — full loops, comp structure, Glassdoor data, failure modes that recur in hiring debriefs. Line all 13 up side by side and a pattern jumps out: every company has at least one round that breaks generic prep. Not a harder version of a LeetCode medium — a structurally different thing that tests a fundamentally different skill.

If you're interviewing in 2026, this is the list you pin to your wall. One quirk per company. The thing that will catch you if you only practiced medium problems in VS Code with the cat on your lap.

---

What each quirk actually tests

Diagram

Copy source

Rendering diagram...

Five non-LeetCode skill clusters explain almost every quirk on this list. Once you see the pattern, your prep gets a lot more targeted. The rest of the post is one quirk per company, with the specific thing to do about it.

---

1. DoorDash — The coding round is an AI session

This is the quirk that started the list.

DoorDash doesn't have a traditional coding round anymore. In its place: a 60-minute AI-assisted working session where you bring your own IDE — Cursor, Claude Code, Copilot, whatever you use daily. You clone a realistic starter project (order dispatch, menu composer, support automation) and extend it (Source: DoorDash careers).

The part that breaks generic prep: "prompt engineering ability" is an explicitly scored dimension. Not a footnote in the rubric — a first-class evaluation criterion. The common failure mode: candidates decide using AI is "cheating" and try to type the whole thing by hand to prove something. They lose signal on every dimension the round actually grades.

What to practice: Spend a weekend building a small project with AI tools in agent mode. Then review every commit you shipped — did you actually read the suggestions, or did you accept them? That review loop is the skill. This isn't coding. It's collaboration under time pressure.

2. Stripe — Bug Squash in a cold codebase

Stripe's most distinctive round is Bug Squash: a live bug bash in a cold codebase. You open an unfamiliar project, read through code you've never seen, find the bugs, and fix them. Finishing 3 of 4 bugs is considered a strong performance.

The real question here is: can you form a hypothesis, write a test that reproduces the bug, and trace back to the root cause — in code you didn't write? That's a completely different skill from "implement two-sum" (Source: Zheng Hao — debugging interview). The candidates who struggle aren't bad coders. They're coders who have only ever practiced writing code, never reading it.

What to practice: Clone an open-source project you've never used. Read it for 30 minutes. Now pick an issue from the tracker and actually fix it. Repeat with a different project every week for a month. You'll stop being scared of cold codebases.

3. Databricks — A full round on concurrency

Most companies skip system design for new grads. Databricks doesn't, and they go further — one full 60-minute onsite round is dedicated to multithreading and race-condition reasoning.

Candidates who aced BFS and DP stall out here because they never had to explain why double-checked locking is broken without volatile. The round goes deep on ConcurrentHashMap internals, thread-safety guarantees, deadlock scenarios, and lock-free data structures. These are topics most CS curricula cover in a single course, and Databricks expects you to reason about them under interview pressure.

What to practice: Java concurrency, specifically: synchronized vs ReentrantLock, volatile semantics, CompletableFuture, and the Java Memory Model's happens-before rules. If you can explain why double-checked locking is broken without volatile, you're in range.

4. Google — You write code in Google Docs

Not a code editor. Not a shared IDE. Google Docs. No syntax highlighting, no autocomplete, no code execution, no red squiggles.

You're still expected to write optimal solutions with clean code — you're just doing it in a word processor. Strong candidates produce code in this round that wouldn't have compiled in any IDE on earth. Muscle memory is a lie when the "Run" button is missing.

What to practice: Open a Google Doc right now. Solve one medium in it. Count the bugs your IDE would have caught. Now solve three more. The strong answer here isn't the clever one — it's the one that compiles on first read.

5. Airbnb — You review PRs instead of writing code

Airbnb added a dedicated Code Review round in 2024. You don't write code. You examine three pre-written pull requests — potentially in languages you don't normally use — identify bugs, suggest improvements, and deliver constructive feedback.

This tests mentoring ability and engineering judgment. Can you read unfamiliar code, spot the real issue (not a style nit), and communicate it like a teammate instead of an auditor? LeetCode has no equivalent for this skill. The strong answer names the trade-off the PR author was making and then the bug — not the other way around (Source: Exponent — Airbnb SWE guide).

What to practice: Start reviewing PRs on GitHub. Pick popular open-source projects and read their recent PRs. For each one, write a review comment identifying the single most impactful issue. Skip the naming-convention nits. Real PR review is about correctness and design — not formatting.

6. Netflix — Being too agreeable is a red flag

Netflix's onsite includes a Dream Team director interview — a behavioral round that explicitly tests keeper-test fit. The memorable detail: being too consensus-oriented or too polite can be a negative signal.

Netflix wants disagreement. Not combativeness — but the ability to hold a position, back it with data, and push back on authority when you believe you're right. Candidates are expected to bring 3+ examples of professional disagreement with real stakes.

What to practice: Prepare 3-5 stories where you disagreed with someone senior and either changed their mind or changed yours based on evidence. The outcome matters less than the process. If all your behavioral stories end with "I just went along with the team," you will not pass this round. Netflix is explicitly allergic to that answer.

7. Amazon — LPs show up inside coding rounds

At most companies, behavioral questions live in one dedicated round. At Amazon, Leadership Principles consume roughly half of total interview time and appear inside coding rounds. Customer Obsession can show up while you're implementing a binary search (Source: aboutamazon.com — using the LPs).

First-loop Amazon candidates routinely get caught by this. LP prep isn't a one-round investment — it's a full-loop investment. Candidates who save their behavioral stories for "the behavioral round" find themselves blanking when the coding interviewer asks "tell me about a time you prioritized the customer" between algorithm questions (Source: aboutamazon.com — an interviewer on the LPs).

What to practice: Prepare 2-3 STAR stories for each of Amazon's 16 Leadership Principles. Yes, all 16. Yes, that's 32-48 stories. The most commonly pulled: Customer Obsession, Ownership, Bias for Action, Dive Deep, and Deliver Results. Tag each story to one principle so you can surface the right one in real time.

8. Meta — An AI coding round with a model chooser

Since October 2025, Meta's onsite includes an AI-enabled coding round in a specialized CoderPad environment — three-panel layout, file explorer on the left, editor in the middle, model chooser on the right with GPT-4o mini, Claude Sonnet/Haiku, Gemini 2.5 Pro, and Llama 4 Maverick as the default (Source: Hello Interview — Meta AI coding). It's rolling out up to E7 and M2.

The real question here is whether you can use AI well. The common failure mode: candidates accept suggestions without reading them — the single most frequent miss on this round (Source: interviewing.io — Meta AI round). The strong answer treats the AI like a fast junior: useful for scaffolding, never trusted without verification. If you rely purely on prompting and can't drive the solution yourself, you fail.

What to practice: Build something non-trivial with AI assistance and force yourself to review every suggestion before accepting it. The workflow is: prompt → read output → find the flaw → refine. The meta-skill is noticing when the AI is wrong before you run the code.

9. Apple — Same week, completely different loops

There is no standardized Apple SWE interview. Two new grads applying the same week to different teams can get completely different loops — different round counts (4 to 6), different question types (some teams skip system design, some don't), different emphasis (domain knowledge for GPU teams, pure algorithms for others).

The common failure mode: candidates prep from Glassdoor posts written by someone on a different team three quarters ago. That person's interview bears zero resemblance to yours because Apple's team-level autonomy means each hiring manager designs their own loop.

What to practice: Ask your recruiter. Seriously. Ask explicitly: "How many rounds are there? Is there a system design round? What should I focus on?" Your recruiter is the single most reliable source of prep information at Apple — more reliable than any prep website. The strong answer here is to not over-prep on the wrong track.

10. Anthropic — The prep packet is part of the test

On the second recruiter screen, Anthropic distributes culture and values documents (Source: Anthropic careers). This isn't background reading — it's homework. Interviewers assess whether you've engaged with the material, and the primary failure mode at Anthropic is mission alignment, not technical ability.

Behavioral red flags that directly disqualify: lone-wolf tendencies, arrogance, financial-only motivation. Rehearsed answers to "why do you care about AI safety?" are detected and penalized. Anthropic wants genuine conviction about safe AI development — and they've built a pipeline specifically designed to tell genuine from performative.

What to practice: Read Anthropic's published research, not the marketing. Form your own opinions. The interviewers are researchers who work on these problems daily — they can tell the difference between someone who read a blog summary and someone who spent an afternoon with the actual papers.

11. OpenAI — Four gates, pass two to pass

OpenAI's coding screens use a four-gate progressive system: you get one custom problem that deepens through four escalating levels. Clear 2 gates and you pass. Finishing all 4 is considered rare.

The quirk: the problem bank is small (~8 known recurring problems). Candidates report seeing the same KV Store, Excel, GPU Credit questions across multiple loops. That makes OpenAI the most "studyable" interview in the StrongYes dossier set — if you can find the questions, you can prepare specifically.

The counter-quirk: the process is widely described as chaotic. Format changes mid-loop, communication gaps are common, and AI tools are strictly prohibited despite OpenAI building GPT-4 and Codex. Practice raw coding with no autocomplete.

What to practice: Seek out the recurring OpenAI problem set (KV Store, Excel, GPU Credit) in public writeups and LeetCode Discuss threads. The bank is small and the questions recur. But prepare to solve them completely from scratch — no IDE, no AI, no hints.

12. Microsoft — A secret round that 70% never see

After the main interview loop, Microsoft runs a separate "As Appropriate" (AA) round with a Director-level interviewer. Only about 30% of candidates reach this stage. The AA interviewer can veto or rescue the hiring decision — overriding the rest of the panel in either direction.

This means your interview isn't truly "over" when you finish the onsite. If the panel is on the fence, the AA round determines the outcome. If the panel recommended hire but the director disagrees, the director wins.

What to practice: Honestly? Nothing specific. The AA round is a senior judgment call about your overall candidacy. But know it exists — so when your recruiter says "we'd like to schedule one more conversation," that's the AA round, and it's a good sign, not an anomaly to be anxious about.

13. Uber — A Bar Raiser can overturn the panel

In the late onsite loop, Uber places a cross-team Bar Raiser who interrogates one past project in detail (Source: interviewing.io — Uber). The Bar Raiser can overturn the rest of the panel — if they say no, the panel's collective yes doesn't matter. This is the same pattern Amazon popularized, and the analysis of why it exists is worth reading if you're interviewing at either (Source: lethain.com — bar raisers and hiring committees).

The memorable requirement: strong answers need hard metrics. Not "I improved performance" — "I reduced p99 latency from 340ms to 85ms, which dropped support tickets by 23% and saved $140K/quarter in infra costs." Uber's Bar Raiser is looking for engineers who measure impact, not describe it.

What to practice: For every project on your resume, prepare hard numbers — latency, throughput, cost savings, user impact, error rates. If you can't quantify it, either find the numbers or prepare a different story. "It worked better" is not an answer. "p95 dropped 4x" is.

---

The pattern

Here's what these 13 quirks share: none of them are about solving algorithms faster.

DoorDash tests AI collaboration. Stripe tests debugging. Databricks tests concurrency. Google tests handwriting code. Airbnb tests code review. Netflix tests disagreement. Amazon tests leadership stories mid-coding-round. Meta tests AI skepticism. Apple tests your ability to handle ambiguity. Anthropic tests values. OpenAI tests deep problem extension. Microsoft has a hidden round. Uber tests measurement.

Generic LeetCode prep covers maybe 40% of what these companies actually evaluate. The other 60% requires company-specific preparation — which is exactly what the dossiers exist for.

Stop prepping for "the interview." Start prepping for their interview.

---

Sources

• DoorDash careers — "Why DoorDash is rebuilding its engineering interviews around AI" · https://careersatdoordash.com/blog/doordash-is-rebuilding-its-engineering-interviews-around-ai/ — DoorDash's own post on the AI Codecraft round and the prompt-engineering rubric dimension.
• Zheng Hao — "Why you should include debugging in the interview process" · https://www.zhenghao.io/posts/debugging-interview — engineer writeup that names the Stripe Bug Squash format and unpacks the debugging-signal dimensions (including the 3-of-4-bugs strong-performance benchmark).
• Exponent — Airbnb SWE Interview Guide (2026) · https://www.tryexponent.com/guides/airbnb-swe-interview — confirms the Airbnb code-review rubric added in 2024.
• Hello Interview — "Meta's AI-Enabled Coding Interview: How to Prepare" · https://www.hellointerview.com/blog/meta-ai-enabled-coding — October 2025 rollout, model dropdown details, CoderPad layout, and drive-the-solution failure mode.
• aboutamazon.com — "How to use Amazon's Leadership Principles in your interview" · https://www.aboutamazon.com/news/workplace/amazon-leadership-principles-interview — Amazon's own 16-LP guide.
• aboutamazon.com — "An Amazon interviewer dives deep into how she uses the Leadership Principles" · https://www.aboutamazon.com/news/workplace/what-do-each-of-amazons-leadership-principles-really-mean — interviewer-side framing.
• lethain.com — "Bar raisers, hiring committees, and other complex ways to improve hiring quality" · https://lethain.com/bar-raising-hiring-committees-hiring-quality/ — staff-engineer-level analysis of why bar raisers exist (used as the Uber Bar Raiser editorial anchor).
• Anthropic careers · https://www.anthropic.com/careers — values-document policy and mission-alignment gate.
• StrongYes dossier store — every per-company quirk traces to the corresponding StrongYes dossier. Interview formats change quickly. If something here is outdated, tell us and we'll update the dossier.