AI Resume Screening: Two Ways to Shortlist Without Guessing
Every HR person I teach has the same week buried somewhere in their calendar. A req closes, 180 applications land in the system, and three people who all have day jobs are supposed to read them, agree on a shortlist, and still sound fair when a hiring manager asks why candidate 41 got a call and candidate 62 did not. Most teams I meet are doing AI resume screening already, in the worst possible way: pasting one resume at a time into a chat window and asking “is this person good?”
That question has no right answer, so you get a confident paragraph of nothing — and a different kind of nothing for each resume, because you phrased it differently at 4:30 than you did at 9:15.
The fix is not a better tool. It is deciding what you are measuring before you start reading, writing it down, and making the AI apply that same thing every time. Here are two ways to do it that you can set up this afternoon.
1. Set Up AI Resume Screening With a Rubric You Write First
Open Claude and start a Project (left sidebar, Projects → New project). A Project is just a folder with a memory: anything you put in its knowledge section is available to every chat inside it, so you are not re-uploading the job description forty times.
Upload two things to the project knowledge: the job description, and a short rubric you write yourself. The rubric is the whole trick:
- Must-haves — 3 to 5 items, each one either present or not. “Has run month-end close unsupervised.” “Holds an active CPA.”
- Nice-to-haves — 3 to 5 items, scored 0, 1 or 2.
- Disqualifiers — the things that end it regardless, like no work authorization for the location.
Then set the project instructions to something close to this, word for word:
For each resume I upload, output a single table row: candidate name, each must-have marked Yes/No/Unclear, each nice-to-have scored 0-2, total score, and one sentence of evidence quoting the resume for every Yes. If the resume does not state something, mark it Unclear. Never infer. Do not recommend hiring or rejecting.
Now drag in ten resumes at a time and you get ten comparable rows instead of ten essays. The two phrases doing the heavy lifting are “mark it Unclear” and “never infer.” Without them the model will cheerfully decide that someone who listed Excel “must know” Power Query. With them, you get an honest pile of Unclears, which is exactly the pile a human should be reading.
The gotcha: that last instruction — do not recommend hiring or rejecting — is not caution for its own sake. Screening is a regulated activity, and a growing number of states and cities put notice, bias-audit and record-keeping obligations on any automated tool that scores or ranks candidates. Keep the AI on “extract and organize what the document says” and keep the decision with a named human. Check with your own IT and legal people about which candidate data is allowed in which tool before you upload a single resume.
2. Turn Interview Notes Into a Scorecard You Can Compare
The second pile is worse, and nobody talks about it. Four interviewers, five candidates, twenty sets of notes — one wrote bullet fragments, one wrote three paragraphs, one wrote “strong!!” and nothing else. Then you sit in a debrief and whoever speaks first anchors the room.
Paste all twenty sets into one chat, in one message, with the interviewers’ names on them, and ask for this:
Build a table: rows = candidates, columns = the competencies we interviewed for. In each cell, summarize what the interviewers actually observed, and after each point put the interviewer's name in brackets. Where two interviewers disagree about the same candidate, put both views in the cell and flag it with DISAGREEMENT. Do not average the views and do not rank the candidates.
What comes back is a grid where the disagreements are visible instead of smoothed away. I have watched a debrief turn around completely because the table showed that the one person who rated a candidate low was also the only one who had asked about handling conflict — which made her worth listening to, not an outlier to be out-voted.
The gotcha: insist on the bracketed names and then spot-check three or four cells against the original notes. Summarizers drift, and the drift is always in the direction of making people sound more agreeable than they were. If a cell has no name attached to it, treat it as the model’s opinion rather than your team’s, and delete it.
Neither of these replaces anyone’s judgment. They just stop you from spending the judgment on reading and reformatting, so you still have some left for the decision.
Plenty more on the calendar
If you want to see this built live rather than read about it, I am running Claude for HR Professionals on Thursday, October 29. We build the project, the rubric, the screening prompts and the debrief table on screen, plus the HR work around hiring — policy questions, handbook lookups, and turning open-ended survey comments into something you can take to leadership. Bring a job description you are actually working on.
If that date does not suit you, the full current schedule is on the front page of PCWebinars.com, and new dates go up as they are confirmed. Register for this one or browse the rest — there is usually something on Excel, Power BI, Copilot or Claude within a week or two.
About your trainer
I’m Tom Fragale. I’ve been a computer trainer and consultant for more than 30 years, and I’ve taught well over 30,000 people how to get more out of the software they already own — Excel, Access, Power BI, Word, Outlook, and now the AI tools that sit on top of all of it: Claude, ChatGPT, and Microsoft Copilot. My background is applied, not theoretical. Alongside teaching, I build real databases and reporting systems for clients in Excel, Access, and Power BI, so the examples I use in class are the problems I ran into last week, not textbook exercises. I also publish free tutorials on YouTube and longer courses online.
You can find the full schedule of upcoming live sessions, along with recordings and course bundles, at PCWebinars.com.