It’s Thursday morning, the job posting you put up on Monday has pulled in 150 applications, and you’re trying to figure out how to sort through them today, on top of everything else you already had planned for the week. A handful are obviously strong, a few are clearly off the mark, and the biggest pile is the in-between: good enough to consider, not yet sorted, and waiting on you.

I’ve been on that side of the desk plenty of times, and most people running small businesses have too. You don’t have an HR team to hand this off to, you don’t have an applicant tracking system running a first pass, and you definitely don’t have a week to read every cover letter and parse every job history.

What you do have is today, a laptop, and a nagging suspicion that somewhere in that stack is the person who would actually change your business — if only you could find them. The good news is that you can, and you don’t need a $500-a-month hiring platform to pull it off. You need a rubric, a screening prompt, and about an hour of focused work.

Before we get into the mechanics, two things are worth getting straight. The rubric is where the real thinking lives; AI is the fast part, but building a clear one takes about thirty honest minutes, and if you skip that step the whole method collapses into guessing at scale. And AI is doing the filtering, not the deciding; this method narrows your stack down to something you can actually read, but the final call on every hire is always yours.

What AI resume screening actually is

Strip away the marketing and all an AI is doing when screening resumes is text comparison. You give it two things: a list of criteria (your rubric) and a resume. It reads both. It compares them. It tells you how well the resume matches.

That’s the whole thing. No sixth sense, no magic. The AI isn’t reading between the lines or picking up the vibe of a candidate. It’s pattern matching on words in front of it.

Which is exactly why the rubric matters so much. The AI will faithfully apply whatever criteria you give it. Vague criteria in, vague scores out. “Good communicator” is vague. “Has written customer-facing copy at a B2B SaaS company for at least two years” is specific. Same AI, totally different results.

The method, step by step

The four steps below are how you turn that pattern-matching capability into a hiring filter you can actually trust. Without structure, asking AI to screen resumes is basically asking it to invent criteria on the fly, which is the vibes-at-scale problem. With a clear rubric out front, you turn the AI into a fast, consistent second pair of eyes that’s working against your real hiring criteria instead of whatever pattern it imagines from a vague prompt.

Each step plays a specific role: define the criteria, hand them to the AI in a way that forces accountability, run the stack, then layer your own judgment on top. The rubric is the load-bearing piece; the rest is scaffolding around it.

Here’s the whole method end to end:

Write the rubric. Decide what makes somebody a fit for this role, written as specific criteria that can be verified straight from a resume.
Write the screening prompt. Use a template that hands the AI your rubric, forces it to quote the resume for every score, and flags one interview question per candidate.
Run the stack. Process resumes in small batches, collect scores in a spreadsheet, and sort to get a ranked shortlist.
Read the shortlist yourself. Layer your judgment on top of the AI’s ranked top 15 to get down to 5-7 candidates worth a phone screen.

The rubric is where we’ll spend the most time, because every other step depends on getting it right. Let’s start there.

Step 1: Write the rubric

Before you open an AI tool, write your rubric. It can be as simple as a list: the criteria you’ll use to evaluate every candidate, split into must-haves and nice-to-haves.

Keep it tight. Five to eight must-haves. Three to five nice-to-haves. Any more than that and you’re basically rewriting your job description into a checklist, and a rubric should do more than that: it should sharpen what actually matters.

The biggest shift to make is this: your criteria should be things a reader can verify straight from the page, not fuzzy traits that only come out in a conversation. If you can look at a resume line and say “yep, that’s evidence” or “nope, that’s not there,” the rubric is doing its job. If every criterion needs a conversation to evaluate, it’s not ready yet, and neither is the AI you’d hand it to.

Here’s what that shift looks like side by side:

Vague rubric:

Team player
Strong communicator
Self-starter
Experience with marketing

Useful rubric:

3+ years managing direct reports
Has written customer-facing copy at a B2B SaaS company
Proficient in HubSpot or Salesforce
Has managed a paid ads budget of at least $10k/month

Same role. Very different rubrics. The first list needs a conversation, a gut feel, and a hiring committee to evaluate. The second list can be checked straight from a resume in about ninety seconds per candidate — which is the whole unlock.

Writing the rubric is also where the real thinking happens. You can’t put together a useful one without deciding what actually matters in this role, and most of the time that’s the hard part. Job descriptions tend to be hedged, inherited, or copy-pasted from the last time somebody hired. Writing the rubric is what finally cuts through that.

Side benefit, and this one sneaks up on you: writing the rubric will tell you whether your job description (JD) is any good. If your rubric asks for things the JD doesn’t mention, or the JD asks for things the rubric doesn’t care about, the JD is broken. Fix it before you screen.

A misaligned JD attracts the wrong applicants, and no amount of clever screening will save you from a bad candidate pool. That’s a lesson I learned the hard way: if your early applicants all seem off, check the JD first. It’s usually the problem.

Build the rubric first. Fix the JD to match. Then start screening.

Step 2: Write the screening prompt

Copy this. Fill in the brackets. Paste into Claude or ChatGPT.

You are screening resumes for a [job title] role at a [company type]. I’m going to paste a rubric and a resume. Score the resume against each rubric item on a 0-3 scale:

0 = no evidence

1 = weak evidence

2 = solid evidence

3 = strong evidence

For each criterion, quote the specific line from the resume that supports your score, or say “not mentioned.”

After scoring, give me:

Total score for must-haves (out of [max])

Total score for nice-to-haves (out of [max])

A two-sentence summary of the candidate’s fit

One flag for anything concerning or worth asking about in an interview

Rubric: [paste rubric]

Resume: [paste resume text]

Two things make this prompt work.

It forces the AI to quote the resume. Every score has to point to a specific line. If the AI can’t find the line, the score is zero. That’s how you catch hallucinations before they turn into hiring mistakes.

It asks for one flag per resume. Every candidate gets a “hmm”: a gap, an inconsistency, a weird pivot, something worth asking about. That flag becomes your interview question. You’re basically doing interview prep while you screen.

What the output looks like

Say you’re hiring a marketing manager. You paste the rubric above and a resume for a candidate named Jordan. What AI sends back looks something like this:

Must-haves:

3+ years managing direct reports: 2. Quote: “Led a team of 2 junior marketers at Acme from 2022-2024”

Written customer-facing copy at B2B SaaS: 3. Quote: “Wrote all outbound email sequences and product launch copy for Beeline, a B2B SaaS startup”

Proficient in HubSpot or Salesforce: 3. Quote: “HubSpot admin for 3 years, built all workflows and attribution reporting”

Paid ads budget of $10k+/month: 1. Quote: “Ran LinkedIn ads on a $3k/month budget” (below threshold)

Nice-to-haves: […]

Summary: Strong SaaS marketing background with solid HubSpot depth and writing skills. Budget experience is smaller than the role requires, which is the main gap.

Flag: Moved from B2C to B2B in 2022. Ask why, and what surprised them about the switch.

Total must-haves: 9/12. Total nice-to-haves: [x/y].

Now multiply that by 150 resumes and you can see why this works. You’re getting a consistent, auditable score with an interview question attached. In about thirty seconds. Per resume.

Step 3: Run the stack

Process in small batches. Three to five resumes per prompt is the sweet spot. You can push to ten, but quality starts dropping as you stuff more into context.

For PDF resumes: Most AI tools take PDF uploads directly now. If yours doesn’t, open the PDF, select all, copy, and paste the text into the chat. Don’t worry about formatting; AI reads the words, not the layout. For scanned PDFs (image files), run them through a free OCR converter like Adobe’s online tool first.

Keep a spreadsheet going. Five columns: candidate name, must-have score, nice-to-have score, summary, flag. Once you’ve run every resume through the prompt, sort by must-have score. The top ten to fifteen names are your shortlist.

Batch smart. Five resumes per prompt, one row per candidate in your spreadsheet, keep moving. The AI doesn’t get tired. You might, but at thirty seconds per resume you won’t have time to.

One more thing that shows up when you do this at scale: AI beats humans on consistency, cold.

You will get tired somewhere around resume 17. Your standards will drift. The candidate at position 150 is not getting the same read as the candidate at position 1. That’s not a moral failing; it’s a fact of being human.

The AI doesn’t have that problem. It applies the same rubric to number 150 that it applied to number 1, and honestly, that alone is worth the setup time.

Step 4: Read the shortlist yourself

AI gave you a ranked top 15. Good. Now do the thing AI can’t.

Read those 15 resumes. Not all 150; the top 15. You’re looking for three things:

Context the AI missed. Career changers. Unconventional paths. Resumes that undersell the person they actually represent. AI scores what’s on the page; you can read between lines.
The flags. Every AI summary came with one. Those are your interview questions. Write them down.
Gut checks. Something pings. Something feels exciting. Something feels off. AI won’t tell you this. Your experience will.

This step isn’t about second-guessing the AI. It’s about adding the judgment layer AI can’t provide. You’re still the hiring manager. AI just cleared your desk so you can actually do the thinking.

Your 15 becomes a 5-to-7 for phone screens. That’s the number a human brain can process well without drift.

About bias (do not skip this part)

AI inherits bias from its training data. This isn’t hypothetical. A 2024 peer-reviewed study from the University of Washington and Brookings audited AI resume screening across nine occupations using 554 public resumes. The findings are hard to read:

White-associated names were preferred in 85.1% of cases; Black-associated names in just 8.6%
Men’s names were favored 51.9% of the time; women’s names only 11.1%
Compared head-to-head against resumes with white men’s names, resumes with Black men’s names were selected 0% of the time

You may have heard of Amazon’s abandoned AI recruiter, the system that penalized resumes containing the word “women’s” and the names of certain all-women colleges. The tool had been trained on a decade of resumes submitted to Amazon, and because tech is a male-dominated field, most of those resumes came from men. Amazon scrapped the project.

The research above shows the underlying bias is still in the models we use today, which means mitigating it is part of the job now.

What to do about it:

Anonymize before you screen. Strip the candidate’s name, address, and any schools that signal gender, race, or other protected characteristics. Replace with “Candidate A, B, C.” One change, huge bias surface eliminated.
Score on criteria, not vibes. The rubric approach already does most of this work. Specific is harder to bias than fuzzy. “Managed a $10k/month ad budget” is harder to discriminate against than “good culture fit.”
Audit the shortlist. After AI returns your top 15, compare the demographic mix to the full applicant pool. If the shortlist looks significantly less diverse, something’s wrong. Re-run with stronger anonymization.
Keep a human in the loop. AI’s output is one input. Your judgment is another. Never automate the final decision.

This isn’t an ethics footnote. It’s a legal issue, a business issue, and a talent issue at once. The candidates you’d miss are often the ones who would change your business the most, and the mitigations above are what help you find them.

The bonus: rubrics pay off twice

The part I wish someone had told me earlier: the rubric you wrote to screen resumes is the same rubric you use in interviews.

When I’m interviewing candidates, I ask permission to record the call. Granola is my go-to for transcription; it runs in the background, captures the whole conversation, and hands me a clean transcript at the end.

After the interview, I paste that transcript into AI along with the same rubric I used for screening, with a simple ask: “Score this candidate against the rubric based on the interview. Quote the specific moments that support each score.”

What comes back is a second opinion, an AI thought partner that scored the same interview I just scored, from a different angle.

Where we agree, I feel more confident in my read. Where we disagree, I go back and re-listen to that stretch of the transcript.

Sometimes the AI caught something I missed. Sometimes I had context it couldn’t hear. Either way, the conversation about the candidate got sharper.

I’m still the one making the decision. But I’m not making it alone, and I’m not making it on a gut feeling that might just be “this person reminded me of someone I liked in 2019.”

The rubric is the thread. Build it once, use it at every stage. Screen, interview, debrief, decide.

This is also the move from Level 1 (chatbox) to Level 2 (skill): turning a one-off prompt into a reusable tool your whole hiring process runs on. You write the rubric prompt once, save it as a Claude skill, and it fires every time you open a hiring conversation. No more setup, no more copy-paste.

More on building custom AI skills for your business here →

What about dedicated AI screening tools?

If you’re hiring constantly (multiple roles a month, always-on recruiting), a dedicated tool might make sense. The dashboards, ATS integrations, and candidate pipelines can be worth paying for if you’re running at that volume.

If you hire a few times a year, it’s probably overhead. The ChatGPT-or-Claude-plus-rubric method works on the free or paid consumer plans you likely already have, and it gives you more control over the criteria. Most of the commercial tools are wrappers on the same underlying AI anyway; you’re paying for the dashboard, not the intelligence.

Where this fits

Resume screening is one of the highest-leverage places to bring AI into a small business. Hiring is expensive. Hiring is easy to screw up. Getting the first filter right saves you dozens of hours and puts better candidates in front of you.

For the bigger picture: How to Use AI to Run Your Small Business →

Related: How to Write SOPs with AI →. Turn your hiring process into a reusable SOP so you’re not rebuilding it every time you hire.

FAQ

Is AI resume screening biased?

Yes, demonstrably. See the Brookings study above. The mitigations in this guide address the main sources of bias, but they don’t eliminate all of it. Stay skeptical, audit your shortlists, and never automate the final call.

How many resumes can I process at once?

Three to five per prompt for best quality. You can push to ten, but fidelity starts slipping. For a stack of 150, expect 30-45 minutes of AI work, plus the thirty minutes you spent writing the rubric.

Do I need a paid AI plan for this?

Free tiers of Claude and ChatGPT handle this fine. The paid consumer plans give you longer context, stronger models, and memory across sessions, which is worth it if you hire regularly.

Can I use AI to write the job description too?

Yes, but write the rubric first. If you write the JD first, you’ll anchor on language that might not match what you actually need. Rubric first. JD second. Screen third.

This guide is for general informational purposes only and does not constitute legal advice. Employment law, anti-discrimination rules, and AI-in-hiring regulations vary by jurisdiction and change often. Before rolling out AI resume screening in your business, talk to an employment attorney licensed in your state about your specific situation.

How to Screen Resumes and Job Applications with AI