Draft every rejection email from the same criteria, not from memory of the interview

Rejecting eight candidates for one role usually means eight separately remembered impressions of eight separate interviews, turned into eight separately worded emails. That's not just inconsistent - it's the kind of inconsistency that looks bad if a candidate ever asks why their rejection sounded more dismissive than someone else's for what was, on paper, a similar outcome.

Feed the model the same scored criteria for every candidate and ask for the rejection email from that, not from a loose memory of how the interview went. If you scored candidates against a set of competencies, give it the scores and ask for a short, warm, specific reason grounded in the actual gap - "the team is looking for more hands-on experience with X at this stage" - rather than either a generic form letter or a candid paragraph about what went wrong in the room.

This does two things at once: every rejection reads at the same standard of courtesy and specificity, and there's a paper trail - the scored criteria - behind every one of them if anyone ever asks why a particular candidate didn't progress. Consistency here isn't a nicety. It's the difference between a decision you can explain and one you're reconstructing from memory months later.