What interviewers score and how levels differ

Rubrics vary from company to company, but most of them collapse into four dimensions. The same prompt is used across levels. What changes is how much the interviewer has to lead you, and how deep you go without being prompted. For the job-side view, see engineering levels.

The four dimensions

How levels differ

• Mid-level (SDE1 to SDE2). A complete, correct end-to-end design is the bar. You cover the functional requirements, your design is reasonable, and you can go deep on one or two areas when the interviewer points at them. Some gaps in breadth are acceptable.
• SDE2 or senior, depending on the company's naming. You find the hard problems yourself and propose deep dives before being asked. Your trade-offs reference the numbers. You bring up failure modes, consistency choices and operations, such as monitoring, backfills and migrations, without prompting.
• SDE3, staff and above. You lead. You reshape the requirements when they are wrong, for example by asking whether likes really need global strong consistency. You explain where the design breaks at the next order of magnitude, and you discuss organizational and evolution costs. You often go deeper than the interviewer planned.

Mistakes that sink strong candidates

• Jumping to components. "We'll need Kafka, Redis and Cassandra" in minute one, before any requirements. Components are answers, and you haven't heard the question yet.
• Buzzword soup. Naming technology without saying what it does here. "Use Kafka" earns nothing. "Use Kafka so fan-out workers can fall behind during spikes without losing events, partitioned by author ID so one author's posts stay ordered" earns a lot. If you can't say what a component does, and what breaks without it, remove it.
• Ignoring the numbers. Computing 175,000 reads a second, then putting one database behind the service. Estimates that never change a decision are a red flag, because they look like ritual.
• No failure modes. Every design assumes everything works. Get there first: replicas, retries with backoff, idempotency, timeouts, and what the user sees during degradation.
• Over-engineering. Multi-region active-active with CRDTs for a system with 1,000 users. Scale the design to the stated requirements, and say what you would add at 10 times the load.
• Going silent. Thinking quietly for two minutes gives the interviewer nothing to score. Think out loud, even when unsure: "I'm deciding between a sorted set and a list here."

Check yourself

1. Why does the same prompt produce different levels? A mid-level candidate completes a correct design. A senior candidate finds the hard problems and goes deep without being asked.

2. Name one signal of a staff-level answer. Reshaping a requirement that is wrong, or explaining where the design breaks at the next order of magnitude.

Key takeaways

• Four dimensions: problem navigation, solution design, technical depth and communication.
• Level is decided mostly in the deep dives, so protect that time.
• The most common sinks are jumping to components, buzzwords without purpose, ignored numbers and missing failure modes.
• Think out loud, and make every choice visible enough to score.

Further learning

Hello Interview's preparation video covers how interviewers at Meta actually score, and what changes by level.

Next, return to the framework and practice the timeline until it is automatic.