Being a good candidate is a skill. Being a good interviewer is a different one: you hold a rubric, you sample work, you take notes that another person could use, and you do not decide in the first two minutes. Unstructured "chat about Kubernetes" hires people who chat. Structured loops still mis-hire, but they fail in ways you can debug.

Before the call

Know the level and the signals (debugging, design, collaboration, domain). Align with the panel so you do not all ask the same system-design question. Send a candidate a realistic agenda. Surprise extra hours of take-home without pay is a filter for desperation, not talent.

A good problem is a slice of the job: a service that times out, a schema that cannot migrate, a PR with a subtle race. Puzzle questions (brain teasers) measure puzzle history.

During

Talk less than you think. Let them think. If they go silent, wait. Hints should be laddered and noted (solved with hint vs without). Do not interview from the answer key in a way that punishes equivalent designs. For behavioral: STAR is a note-taking aid, not a ritual humiliation. Probe their decisions, not the team's brand names.

Write evidence: "Proposed outbox for dual-write, named the failure if consumer is not idempotent." Not: "Seemed strong."

After

Score against the rubric before the debrief so the loudest staff engineer does not rewrite everyone's memory. A "no" needs a dimension. A "yes" needs risks (needs mentoring in X). Shadow a calibrated interviewer. If your yes-rate is 90% or 5%, you are not using the same bar as the panel.

Legal and ethics: avoid questions that elicit protected information. Stick to the job.

Read a structured hiring chapter (Smart/Street or re:Work). Then rewrite one question you always ask so it has a rubric of 1–4 with examples. The loop is a measurement instrument. If you cannot say what you measured, you measured confidence. Confidence is not a level.