The Interview Loop, Stage by Stage

The loop turns the pillars into a sequence of conversations, each producing a specific signal, arranged so the cheapest gates come first. The design rule that makes it maintainable: every round is a standalone document containing its own purpose, timing, probes, example questions, and rubric rows. An interviewer should never need a second tab open.

Thirty minutes, run by a recruiter, graded on concreteness, not technical depth. It reads four signals — skin in the game, a belief the person has actually changed, a subtraction they made, and something they built from zero — and it clears the hard gates early so no one wastes a panel. The arc: framing (0–5), motivation and logistics including the inversion question "what would make this the wrong fit in six months?" (5–10), built/owned/AI background with one named moment per answer (10–20), two of the four signal probes (20–25), loop preview and close (25–30). Buzzword answers get restated and re-asked once, then down-scored.

Thirty minutes with the person's future manager. If this is soft, the panel rarely saves it. It checks culture fit, technical breadth, the ability to operate autonomously, and a real zero-to-one story. Three asks are stated to the candidate up front: be concrete (scope, stakes, who decided, what shipped); use "I," not "we"; and push back when you think I'm wrong.

The round mixes a sharp background pass, a rapid-fire AI segment (thirty to sixty seconds each: the dumbest thing you let AI automate this year, your agentic setup, the most overrated tool, the one thing you would not trust AI to do in production), an autonomy-and-strategy segment, and a short team/leadership segment. It ends on one deliberately hard fundamentals question — for a platform role, a good one is "What does high availability mean to you, and how could you fake it?" You are listening for the vocabulary of redundancy, failover, and load-balancing; for the distinction between availability and disaster recovery; and for signal-versus-noise reasoning. A candidate who cannot separate real redundancy from a green dashboard has a shaky foundation.

Sixty minutes with a senior engineer, and the primary go/no-go. About twenty-five minutes are a live diagnostic on a shared whiteboard using real artifacts — configuration, command output, event logs, dashboards — never "imagine a pod." The rest probes mentality, with a short wrap.

The diagnostic is request-driven: seed only the trigger and a reassuring-but-false green dashboard, then make the candidate ask for each artifact by name. What they ask for, and in what order, is the grade. Build it in escalating parts so a fast candidate never runs out of road and a slower one is not penalized for spending the whole window on part one.

A worked example: intermittent timeouts after a routine deploy, dashboard still green. Root cause — the new revision raised application concurrency but left the database connection pool small, so requests queue and fail at the acquire timeout, while fast cache reads keep the average panel green even as p99 is pegged. The scenario tests layered debugging, reading evidence before acting, spotting both the lying dashboard and a health check that never touches the database, restraint (restart / scale / bump-the-pool are motion that mask the cause), and a sized fix that respects the database's connection ceiling. Restarting or scaling before reading anything is the clearest junior tell there is.

Sixty minutes, drawn live on a shared canvas. Grade the quality of inquiry, isolation and blast-radius thinking, failure-mode awareness, and production-readiness judgment — not architecture recall. The candidate picks a system and must ask clarifying questions before designing; designing without asking scores down. Push on the seams: where do builds run and what's the blast radius of a malicious build step; one tenant floods the queue, what happens to everyone else; the rollback story; where secrets live during a build. Score each section against its own level ladder and take the modal level as the overall, with architecture and failure-modes as tiebreakers.

The values / cross-functional panel (45m, mixed panel) checks how the person operates at the seams: absorbing cost to unblock a partner team, holding the line versus folding, teaching a non-specialist team something. The executive round (30m) checks ambition, taste, and fit: a first-year bet they'd want to make and its stakes, the worst decision they've seen in the domain, why here and why now.

At the end of every round, the questions the candidate asks are scored. Strong questions probe decision rights, what the team has removed recently, where the team admits it got something wrong, and what would make this role miserable. Weak questions are only about tooling, perks, and comp. What a person is curious about reveals how they'll operate once inside.