Create and Apply a Resume Screening Rubric

Workflow area: HR & Recruiting · Tier: Core · Hands-on time: ~35 minutes

The rubric matters more than the AI. A written rubric forces you to say, before seeing any candidate, what evidence counts and how much — which is the same discipline that makes structured interviews (pb-30) work. AI's role here is consistency at volume: it applies your rubric the same way to resume #7 and resume #107. What it must never become is the decider; in several jurisdictions, automated employment screening carries specific legal obligations, and everywhere it carries fairness risk.

When to use — and when not to

• Use it when volume breaks human attentionmore than ~30 resumes for one role is where consistency collapses without a rubric.
• Use it to make screening auditablea scored rubric per candidate is a defensible record of why decisions went the way they did.
• Use it when multiple people split the screening pileshared rubric, comparable scores.
• Skip fully automated rejectionno candidate is rejected because a model scored them low. The ranking orders your reading; a human makes every call. This is a legal requirement in some places and good practice everywhere.
• Skip criteria that proxy for protected characteristicsgraduation years (age), "native English" (origin), unbroken employment history (parental leave, illness). If the rubric encodes bias, AI applies that bias with perfect consistency.

The recipe

1. 1.

   Draft the rubric from the JD's essentials. Use the ESSENTIAL requirements from your JD work (pb-29):

   Copy

   Build a resume screening rubric for this role.
   
   Role: [TITLE, LEVEL]
   Essential requirements: [LIST — from the JD]
   Nice-to-haves: [LIST]
   What the best performer in this role had on day one: [1-2 LINES]
   
   Produce a rubric with:
   1. 4-6 weighted criteria (weights sum to 100), each tied to an
      essential requirement
   2. For each criterion, a 0-3 scale where each level names the
      RESUME EVIDENCE that earns it (e.g. "3 = led reconciliation
      for multi-entity books; 1 = mentions Tally/Excel with no
      outcome stated")
   3. Two KNOCKOUTS only if truly non-negotiable (e.g. work
      authorization) — flag, not score
   4. A BIAS CHECK: list ways this rubric could disadvantage
      candidates with equivalent ability (career breaks,
      non-traditional paths, non-brand employers) and adjust the
      evidence descriptions to credit equivalents.
   
   Rules: criteria must be assessable from a resume alone — no
   personality guesses. No criteria touching age, gender, family,
   nationality, religion, health, or proxies for them.
   

2. 2.

   Calibrate the rubric on three known resumes. Before the real pile, test on resumes where you know the answer — your current strong performer, a past hire who struggled, a past reject. If the rubric ranks them wrong, fix the weights or evidence descriptions now. This ten-minute step is what separates a rubric from a wish list.

3. 3.

   Redact, then apply in batches. Strip names, emails, addresses, photos, and graduation years from the resumes (many ATS tools export anonymized versions; otherwise find-and-replace with Candidate-01, Candidate-02…). Then:

   Copy

   Apply this rubric to the following resumes. For EACH candidate
   output a table row: candidate ID | score per criterion | weighted
   total | knockout flags | 2-line evidence summary QUOTING the
   resume lines that earned the top two scores | "UNCERTAIN" flag if
   the resume is ambiguous on any criterion.
   
   Rubric: [PASTE FINAL RUBRIC]
   
   Resumes:
   [PASTE 5-10 REDACTED RESUMES PER BATCH, SEPARATED BY ===]
   
   Rules: score only from stated evidence — do not infer skills from
   job titles alone. When evidence is absent, score 0 and mark
   UNCERTAIN rather than guessing. Do not reward resume polish,
   formatting, or prestigious employer names.
   

4. 4.

   Verify before you trust the ranking. A usable output row looks like:

   Copy

   Candidate-07 | Ledger ops 3, Reporting 2, Systems 2, Comms 1 |
   Total 71/100 | No knockouts | Evidence: "owned month-end close
   for 3 entities, cut close time 9→5 days"; "built AR aging
   dashboard in Sheets" | UNCERTAIN: communication — no evidence
   either way
   

   Spot-check 20%: open three scored resumes and confirm the quoted evidence exists and the scores follow the anchors. If quotes don't match the resume text, the batch is unreliable — re-run smaller batches.

5. 5.

   Human decisions, with the bias check built in. Read every resume above your line and every UNCERTAIN flag. Then run the standing check: compare the advance list against the full pool for patterns (all from the same three companies? no career-break candidates?). Ask the model: "Here are the score distributions — what systematic pattern should a human review before finalizing?" The final advance/reject call — and the accountability for it — is yours, recorded per candidate in your sheet.

Make it yours

• Google Sheets as the system of record: paste batch outputs into a sheet — one row per candidate, plus columns for "human review done", "decision", "decision reason". That sheet is your audit trail.
• Two-model cross-check: for senior or contentious roles, run the same batch through Claude and ChatGPT and flag any candidate whose totals differ by more than 15 points — disagreement means ambiguous evidence, which means a human reads closely.
• Rubric reuse: the rubric (minus role specifics) becomes a template; next opening, you update criteria and re-calibrate on known resumes in ten minutes.

Quality check

• The rubric was calibrated on known outcomesit ranked your known-good and known-poor resumes correctly before touching the live pile.
• Resumes were redacted before pastingno names, contact details, photos, or graduation years reached the model.
• Every score carries quoted evidenceand spot-checks confirmed the quotes are real.
• No candidate was rejected by a number alonea human reviewed every decision, including UNCERTAINs.
• The pattern check ranyou looked at who the process filters out, not just who it ranks up.