Technical Interviews That Actually Predict Performance
Most technical loops measure interview skill. A few measure the job. The difference is in what you ask candidates to do.
We have run several thousand technical assessments and tracked the ones that survived contact with the job. The formats that predicted performance have almost nothing in common with the formats that are most common.
What did not predict anything
- Algorithm puzzles unrelated to the team's actual work
- Take-home exercises longer than three hours
- Trivia about framework internals
- Whiteboard coding without an execution environment
What did
The single strongest format was a short, realistic problem in the candidate's own environment, followed by a conversation about the trade-offs they made. The code matters less than whether they can explain what they chose not to do.
- 01Give a problem drawn from a real ticket, reduced to sixty minutes.
- 02Let them use their own tooling, including an AI assistant.
- 03Spend half the session on the reasoning, not the output.
- 04Ask what they would change with another week.
- 05Score against a written rubric agreed before the first candidate.
“You are not hiring someone to write code alone in a room. Stop assessing them as if you were.”
Calibration is the unglamorous half
A rubric written after the first three candidates is a description of those candidates. Agree the bar in writing before the loop opens, and have two interviewers score the same session independently at least once per search.
