Writing a screening bar candidates can pass fairly

A must-have is a promise: any candidate missing this will be declined, no matter what else they bring. Write each one knowing it will be enforced literally and consistently — that's what you're asking the agent to do. If you're not willing to lose a great candidate over a criterion, it isn't a must-have.

The fairness test: a must-have should be observable in an application. "Has run experiments on a production ML system" is observable. "Passionate about our mission" is not — no applicant can evidence it and no screener can check it, so it degrades into noise that penalizes people who write plainly. Keep must-haves to the three or four things the first year of the job genuinely cannot work without.

Nice-to-haves rank the candidates who cleared the bar; they must never reject. The failure mode is a nice-to-have doing silent must-have work — "bonus: fintech experience" that in practice sinks everyone without it. Ranking should degrade gracefully: a candidate with zero nice-to-haves who clears every must-have is, by your own definition, worth an interview.

This split is also where diversity of background survives or dies. Career-changers, bootcamp grads, and people from adjacent industries usually clear well-written must-haves and lose on pattern-matching nice-to-haves. Keeping the two ledgers separate — and making rejection impossible on the second ledger — is the single highest-leverage fairness decision in the whole pipeline.

An honest screener shows its work. Every Tandem screening decision cites evidence: this requirement, met by this line of the application, or missing entirely. That audit trail isn't decoration — it's what lets you distinguish "the agent misread" from "my bar was wrong," which are fixed in different places.

Review your declines, not just your advances. A half-hour skimming declined candidates with their cited reasons will surface a miscalibrated criterion faster than any dashboard. When you find one, fix the plan and the fix applies to every future applicant — which is precisely the leverage a human screening process never gives you.

Weak bar: "5+ years backend experience, strong communicator, startup mindset, Go preferred." Every clause is either a proxy (years), unobservable (mindset), or a nice-to-have cosplaying as a requirement (Go).

Fair bar — must-haves: has built and operated a production API service; has owned a data model through a real migration; writing sample or application answers show clear technical reasoning. Nice-to-haves: Go, experience in a sub-30-person team, marketplace domain. Same role, but this version screens in every capable person the weak version would have bounced — and the agent enforcing it can tell you exactly why each decision went the way it did.