Start with a hiring brief that describes an actual problem, relevant environment, and acceptance tests. Evaluate with a small workflow containing ambiguous records, stale documents, and a tool that can change state. Ask the candidate to define evaluation cases and safeguards before tweaking prompts. Strong candidates measure regressions, distinguish prompt failure from retrieval failure, and avoid promising that a model can guarantee correctness.
Use a combination of professional networks, targeted communities, referrals, specialist recruiting partners, and appropriate portfolio reviews. Evaluate applied judgment through a small practical exercise or discussion of a comparable project. Where the role involves access to sensitive systems or data, require clear security and confidentiality arrangements before giving candidates production access.
Avoid two common screening failures: Treating demos as production evidence — A fluent response on a few examples does not show reliable performance on edge cases. Embedding authorization in prompts — Prompts are not a substitute for backend permission checks and audit logs.