Screen CVs Fairly

You turn a pile of applications into a shortlist that a hiring manager can trust and defend. The standard: the same criteria, applied the same way, to every applicant, with written evidence for every score, and no use of protected characteristics or their proxies. You recommend; a human decides. Screening is where most unfairness in hiring happens quietly, so the process matters as much as the result.

When to use

• "We've had 140 applications for the ops role, give me a shortlist."
• "Go through these CVs and tell me who to interview."
• "Is this candidate worth a call?"
• "Draft the rejections for everyone we're not progressing."

What you need

• The must-haves and nice-to-haves for the role, agreed with the hiring manager before you open a single CV. If they do not exist, stop and get them (see write-job-description).
• The applications. Access is a connected app (the owner's applicant tracking system) or the agent's own browser signed in by the owner, or files the owner gives you. If none, ask with clarify.
• The owner's privacy notice and retention period for applicant data.
• The countries where applicants live and where the job is. Discrimination and data law differ; check the local rules with web_search and record sources.

Method

1. Fix the rubric first. Write screening-rubric.md using references/screening-rubric.md: each must-have becomes a criterion with a 0 to 3 anchored scale and examples of evidence at each level.
   • Nice-to-haves get a lighter weight or are used only to break ties.
   • Get the hiring manager to approve it with ask_teammate.
   • Lock it: changing criteria halfway is how bias gets in.
2. Separate the identity data. Use execute_code to make a redacted copy of each application for scoring.
   • Remove or mask: name, photo, date of birth and age, gender markers and pronouns, marital and family status, nationality and place of birth, home address beyond the city or country relevant to the role, graduation and school-leaving years, and memberships that reveal religion, ethnicity, politics or health.
   • Key each copy by an applicant ID.
   • Keep the ID-to-name map in a separate private file.
3. Calibrate. Score the first 10 applications, then ask the hiring manager to score the same 10 blind.
   • Where any criterion differs by more than one point, discuss the anchor and rewrite it before you continue.
   • Re-score the 10 if the anchors changed.
4. Score every application the same way. For each criterion, give 0 to 3 and quote or point to the exact evidence.
   • "No evidence found" is a 0 only for that criterion; it is not a judgement of the person.
   • Do not reward polish, design, brand-name employers or prestigious universities unless a criterion asks for them.
   • Score in batches of about 20 and re-read your first and last score of each batch to check you have not drifted.
5. Apply the decision rule. Default rule (adjust with the hiring manager): progress if every must-have scores 1 or more and the weighted total is at or above the agreed threshold.
   • Mark as borderline anyone within one point of the threshold, anyone with a single must-have at 0 but strong elsewhere, and anyone whose application is unusual (career change, non-standard education, gaps, overseas qualifications).
   • Borderline cases always go to a human; you never reject them yourself.
6. Handle the hard cases fairly. Career gaps: ignore unless the role has a lawful, job-related recency need; gaps often reflect caring, illness, disability or study.
   • Overseas qualifications: judge the skill, and if a formal equivalence matters, flag it for checking.
   • Right to work: check only whether the person states they have it or need sponsorship, if the owner's process asks, and treat everyone the same way.
   • Adjustment requests: route them to the owner straight away; they are never a screening factor.
7. Remember the law on automated decisions. In the EU, a decision based solely on automated processing that significantly affects someone is restricted (GDPR Article 22). In the UK, Articles 22A to 22D of the UK GDPR replaced Article 22 on 5 February 2026: such decisions are allowed only with safeguards (tell the person, let them make representations and contest it, offer human review), with tighter limits where special category data is used. AI used to screen or rank applicants is classed as high-risk under the EU AI Act (Annex III, point 4(a)); after the 2026 Digital Omnibus on AI, the high-risk obligations for these systems apply from 2 December 2027, so check which obligations apply on the date you work.
   • Some places, such as New York City, require bias audits of automated hiring tools.
   • So: you score and explain; a named human reviews and decides every outcome.
   • Say this in your hand-over.
8. Check for adverse impact (only if the owner holds monitoring data). Never look at demographic data while scoring.
   • If the owner's ATS collects voluntary equal opportunity data, after the decisions a human (or you, with the owner's explicit permission, on aggregate data only) can compare selection rates by group.
   • The four-fifths rule from the US Uniform Guidelines (29 CFR 1607.4(D)) is a common rough test: a group's selection rate below 80 per cent of the highest group's rate is a flag to investigate.
   • With small numbers it is unreliable; say so.
   • See the calculation in references/screening-rubric.md.
9. Draft the communications. For applicants not progressing, draft respectful, prompt rejections with references/rejection-drafts.md.
   • For applicants progressing, draft the next-step invitation.
   • Nothing is sent until the owner approves.
10. Hand over. Use show_card for the shortlist and the borderline list, with scores and evidence. Store the scored sheet where the owner keeps hiring records, not in chat.

Judgement calls

• No cover letter: do not penalise it unless the ad asked every applicant to answer specific questions, and then score only those answers.
• The same person applied for two roles: score each application against that role's rubric, separately.
• Internal applicants: same rubric, same process. Their internal status matters only for process steps the policy sets (such as telling their manager), never for the score.
• Too many applications to score carefully in the time given: tell the owner instead of skimming. Options are more time, a second scorer, or short knock-out questions approved by the hiring manager and applied to everyone.
• A strong applicant who misses one must-have narrowly: borderline, with a note on what an interview would need to test.

Output

• screening-<role>.csv: applicant ID, score per criterion, evidence per criterion, weighted total, recommendation (progress, borderline, do not progress), reviewer notes.
• A shortlist card: top candidates by ID with one-line evidence summaries; a separate borderline card with the reason each is borderline.
• rejections-<role>.md: draft messages, grouped by stage, awaiting approval.
• A short method note: rubric version, calibration result, threshold, who must make the final decisions.

Checks before you finish

• The rubric was approved before scoring and did not change during scoring (or, if it did, everyone was re-scored).
• Every score has evidence; no score is based on the name, photo, age, school prestige or formatting.
• Borderline and unusual applications are flagged for a human, not rejected.
• The identity map is stored separately and privately; nothing personal is pasted into shared channels.
• No message has been sent. Rejections and invitations are drafts for the owner.
• The hand-over names the human who will make the decisions.

Pitfalls

• Criteria creep. "Oh, and they should know Salesforce" halfway through. Stop, agree it, and re-score everyone or drop it.
• Pedigree bias. Big-name employers and universities are not evidence of the must-have. Read what the person did.
• Keyword knockouts. A CV that says "ledger reconciliations" instead of "account reconciliations" still has the skill. Read for meaning.
• Penalising gaps or part-time work. These correlate with sex, disability and caring responsibilities. Do not score them.
• Speed over consistency. The last 30 CVs get less attention than the first 30. Batch and re-check.
• Silent rejection. Not replying is the worst candidate experience. Draft every rejection.

See also: structured-interview-kit for the next stage and write-job-description for the must-haves.