Skip to content
Hiring teams7 min read

A Resume Scoring Rubric That Works

A resume scoring rubric for hiring teams: anchored levels, evidence rules, knockouts, blind review and calibration. Includes a free template.

By the HireRabbit.AI team · Published · Last updated

Reviewers mark a clipboard rubric with checkboxes and star ratings beside resumes under a magnifying glass
On this page
  1. Why a written rubric beats a good reader
  2. Score what people did, not years and credentials
  3. Keep knockouts separate and few
  4. Write anchored levels for each criterion
  5. Gaps, career changers and non-traditional backgrounds
  6. Hide names and photos before anyone scores
  7. Calibrate reviewers before they score the real pile
  8. Free resume scoring rubric template: support engineer
  9. How this maps to HireRabbit.AI

Most hiring teams score resumes, but few write down how. Two recruiters read the same application. One sees a strong support engineer and the other sees someone who changes jobs too often, and neither can point to the line that decided it.

This is a resume scoring rubric for employers, not the student kind that grades layout and grammar. It covers what to score, knockouts, anchored levels, gaps, bias and calibration, and it ends with a free template for a customer support engineer.

Why a written rubric beats a good reader

Kuncel and colleagues' 2013 meta-analysis in the Journal of Applied Psychology compared an expert weighing applicant information in their head with a fixed formula. For predicting job performance, mechanical combination improved prediction by more than 50%, even when the experts knew the job and the organization well.

Expertise helps decide what to look for, but not to add up several judgments the same way on the fortieth resume as on the first. A rubric makes the result depend far less on who reads.

Score what people did, not years and credentials

Counting is the weakest way to read a resume. A 1988 review by McDaniel, Schmidt and Hunter found that awarding points for years and credentials has little predictive support. Scoring described achievements against anchored scales was the strongest of the four methods they compared.

Years of experience do no better. In a 2019 Personnel Psychology analysis of 81 samples, Van Iddekinge and colleagues found that pre-hire work experience correlated .06 with job performance, where 0 means no relationship and 1 a perfect one.

So score the content of experience, not its length. "Five years in support" says how long someone held a job. "Reproduced customer-reported API errors from logs and wrote the fix steps for the help center" says what they can do.

Keep knockouts separate and few

A knockout is a pass-or-fail check that runs before scoring, often a yes-or-no question on the application form. Keep knockouts to true minimum qualifications tied to the job, and write down why each one is there.

In Griggs v. Duke Power (1971), the US Supreme Court held that requirements such as a diploma that exclude protected groups at a higher rate must be related to the job.

Keep two to four knockouts per role and give no points for passing. Have a person confirm every rejection, because a misread question can knock out someone whose resume shows the qualification. Sorting must-haves from weighted requirements is covered in the job description guide.

Write anchored levels for each criterion

Four criteria cover most roles: skills, experience, projects and education. The explainable AI resume screening guide introduced them with a 0 to 100 range. Each needs levels that describe evidence a reviewer can point to.

A 0 to 3 scale works well at the resume stage. There is no middle point to retreat to, and 0 means "no evidence", so a missing section is never confused with a weak one.

Teams that use 1 to 5 elsewhere, as in the structured interview scorecard guide, can use it here and compute the overall as (weighted sum minus 100) divided by 4.

Put the evidence rule at the top of the scoring sheet. A skill tied to a task counts, such as "Queried Postgres to confirm which accounts hit the billing bug," and so does work done anywhere, including volunteering, courses and personal projects. Adjectives, claims nobody could check and proxies such as school prestige or years in a role do not.

Gaps, career changers and non-traditional backgrounds

Write a policy for these before you open the first resume. In a 2013 field experiment across 100 US cities, Kroft, Lange and Notowidigdo found that callbacks fell as the unemployment spell on a resume grew longer.

So gaps are not scored or mentioned in review notes. Volunteer work, freelance work, courses and projects are scored on the same anchors as paid roles, and a bootcamp or certificate counts the same as a degree where it covers the same ground.

A February 2024 report from the Burning Glass Institute and Harvard Business School found that 37% of firms had removed degree requirements, yet the change showed up in "not even 1 in 700 hires" in 2023. Dropping a requirement from the posting changes little if reviewers still give points for the degree.

Hide names and photos before anyone scores

In Bertrand and Mullainathan's 2004 study, fictitious resumes with white-sounding names received 50% more callbacks. Kline, Rose and Walters' 2022 study of more than 83,000 fictitious applications found distinctively Black names lowered the chance of contact by 2.1 percentage points.

Software does not remove the problem on its own. In a 2024 University of Washington study, three large language models preferred white-associated names 85% of the time.

Whether a person or a tool does the first read, check pass rates by group afterwards at the knockout step and the score cut-off, using the method in the adverse impact analysis guide.

Before anyone scores, remove the name, photo, home address, email handle and graduation years. None of them is evidence for any criterion.

A resume with covers over the name, photo, home address, email handle and graduation year, none of which is evidence for any criterion. Point at a cover to peel it up, or at an open block to see what gets scored.

Calibrate reviewers before they score the real pile

A rubric that two reviewers read differently is still two rubrics. In the US Department of the Interior's Structured Resume Review Instructions, a panel independently scores each competency 1 to 5, panelists more than one point apart discuss it, and hiring managers cannot sit on the panel.

We recommend the same for a private team, with a one-line evidence note on every score.

A 30-minute calibration routine

Once per role, take five redacted resumes from past applicants: one obvious no, one obvious yes and three in between.

  1. Minutes 0 to 5: read the criteria, anchors and evidence rule aloud.
  2. Minutes 5 to 20: each reviewer scores all five alone, with evidence notes.
  3. Minutes 20 to 27: discuss only cells where scores differ and agree which anchor the evidence fits.
  4. Minutes 27 to 30: rewrite any anchor that caused disagreement, and date the change.

Run it again when the role changes or a new reviewer joins.

Free resume scoring rubric template: support engineer

This template is for a customer support engineer at a 200-person software company, handling technical escalations from business customers. The must-haves are debugging customer-reported issues, reading logs and basic SQL, and writing for non-technical readers.

CriterionWeightLevel anchorsEvidence examples
Skills40%0: none of the three must-haves appear. 1: must-haves or tools listed in a skills section only. 2: one or two must-haves tied to specific work or projects. 3: all three tied to specific work, plus a help-desk tool or web APIs and Postgres named in a role or project."Traced failed webhook deliveries in application logs"; Zendesk or Intercom named in a role
Experience30%0: no work involving software or customer problems. 1: customer-facing work outside software, or software work with no customer contact. 2: handled customer-reported technical issues in a support, QA or engineering role. 3: owned escalations end to end, from reproducing the issue through the fix to the customer update.Ticket queues worked, escalations owned, bug reports filed with reproduction steps
Projects20%0: no projects, or names with no description. 1: a described project unrelated to troubleshooting, data or documentation. 2: a related project with the candidate's own part stated. 3: a project that shows debugging or explaining, with the candidate's part and a result stated.A small app built on a web API and Postgres; a published runbook
Education10%0: no related education or training listed. 1: related coursework or a short course. 2: a completed related program, whether degree, diploma, bootcamp or certificate. 3: a related program plus a credential the role names, such as a certificate in SQL. Score a capstone under Projects, not here.Ignore school names, grades and graduation years

Overall score = (40 × skills + 30 × experience + 20 × projects + 10 × education) ÷ 3, a number from 0 to 100. A candidate scored 2, 3, 1 and 2 gets (80 + 90 + 20 + 20) ÷ 3 = 70. Put the formula in a spreadsheet: that fixed step is the mechanical combination Kuncel's meta-analysis supports.

Each lane is as wide as its weight, 40, 30, 20 and 10, and holds one block per level scored, so the overall score is how full the tray is: 2, 3, 1 and 2 fill 70. Point at a cell to change that criterion's level and watch the score in the corner.

Before scoring, check the knockouts: stated hours and location, and at least one of the three must-haves anywhere on the resume, including projects and courses. A resume with none goes to a person for a second look.

Record the knockout result, four level scores with evidence notes, the overall score, the dated rubric version, the reviewer and the decision with a written reason. The guide to documenting hiring decisions goes further on building that record.

How this maps to HireRabbit.AI

HireRabbit.AI follows the same structure, with the AI doing the first read. Every resume gets four area scores, and the overall score is computed from those four using weights the AI sets from the job description, which stay the same for every applicant.

Auto-reject runs only below the cut-off your team sets, in a batch a person confirms, and it can be undone. The how scoring works page shows the mechanics.

Put this into practice.

Free during invite-only early access. All you need is a work email, and a person reviews every request.