AI-first reference check script

AI-First Reference Check

Purpose

Validate the leverage and judgement signals from the interview loop with someone who watched the candidate work every day. A reference call is not a box-tick — it is the last chance to catch a Lazy User or confirm an Architect before the offer.

Structure (15 minutes)

1. Role context (2 min). What was the candidate's core outcome, and how hands-on were you with their work?
2. Leverage probe (4 min). "Can you describe a task the candidate used to do manually that they stopped doing? What replaced it, and did it hold up?" Listen for whether the systematisation actually lasted.
3. Judgment probe (3 min). "Where did you see them push back on an AI output or catch an error no one else did?" No example = no Auditor signal.
4. Sharing probe (3 min). "Did their work make anyone else faster? How?" Confirms the Multiplier pillar — or reveals gatekeeping.
5. The honest question (3 min). "If you had to name one situation where they trusted a tool they shouldn't have, or over-engineered something simple — what was it?"

Output format

Return:

• Recommendation: confirm offer / hold / flag concern.
• Pillar confirmation: for each of Architect, Operator, Auditor, Multiplier — confirmed / not seen / contradicted, with one line of evidence.
• Strongest quote: one verbatim line from the reference.
• Risk: the single biggest open question, and a suggested question for the hiring manager to resolve it.

Rules

• Never accept "they were great, no notes." Push for a specific story on every probe — generalities are not evidence.
• A contradiction between the reference and the interview loop is the most important signal in the call. Surface it explicitly.
• If the reference cannot name a single tool the candidate systematised away, downgrade the Architect score from the interview.
• Do not score a pillar the reference never actually observed — mark it "not seen" rather than guessing.