How to automate candidate screening with AI
All 180 applications scored against the same written rubric by 8am, each with the deciding signal named, and the notes already on the candidate records.
Application 12 gets read generously at 9am. Application 140 gets four seconds at 6pm. That is fatigue rather than merit deciding who you interview.
A companion reads all 180 against a rubric you wrote down, scores each one, names the single signal that decided it, and writes the note back onto the candidate record. Stage moves and rejections stay with you. Finding people who never applied is the candidate sourcing playbook.
| Read applications and candidates | Yes | Yes | Yes | Yes |
|---|---|---|---|---|
| Read interviewer scorecards | Yes | Not exposed | Not exposed | Not exposed |
| Write a note back | Candidate activity feed | Opportunity note | Attachment only | Candidate comment |
| Move a stage | Human, in the UI | Human, in the UI | Yes, with approval | Yes, with approval |
| Reject a candidate | Human | Human | Human | Human |
Turn the job post into a scoring rubric
A job description is a marketing document.
A rubric is a decision procedure, and writing the second one is where inconsistent screening stops.
Prompt: “Read the job description for our Senior Data Engineer role in Greenhouse and turn it into a screening rubric. Propose five weighted criteria with what a 1, 3, and 5 looks like for each, and list the hard disqualifiers separately. Where the job post is vague, it says ‘strong SQL’ without saying what that means, ask me instead of deciding. Print the rubric so I can edit it before we use it.”
The questions that come back are the useful part. Most posts carry two or three criteria nobody had agreed on.
How do you read 180 applications without drifting?
A companion does not get tired at application 140.
Same rubric, same evidence standard, all the way down the pile.
Prompt: “Pull every application for that job in Greenhouse with the candidate’s attached resume. Score each against the rubric, one row per candidate: name, application id, score per criterion, total, the single signal that decided the score, and any disqualifier hit. Where the resume is ambiguous about a criterion, score it as unknown rather than guessing, and list those separately. Write it to a sheet, sorted by total.”
Then audit the middle band: “Show me the fifteen candidates scoring in the middle band and the exact resume line you scored each on.” A correction in plain language, “a bootcamp plus two years shipping production pipelines counts as equivalent”, carries into every later run.
Where does the write-back stop?
The four common ATS platforms expose different surfaces, so what happens after scoring depends on which one you use.
Prompt: “For every candidate scoring 4 or above, post a note on their Greenhouse activity feed with the score, the deciding signal, and the two questions a screener should ask them. Do not change anyone’s stage and do not reject anyone. I will do both in the ATS.”
On Ashby or Workable you can hand the stage move over too: “Move everyone I have ticked to Phone Screen in Ashby, and show me the list before you do it.” On Greenhouse and Lever advancement happens in the ATS, and rejection stays with a person everywhere.
Score new applications every morning
Applications arrive continuously and screening happens in bursts, which is how a good candidate waits nine days for a first reply.
Prompt: “Every weekday at 8am, score any application added to our open roles since the last run using each role’s rubric, write the notes back for anyone scoring 4 or above, and Slack me a summary: how many new, how many strong, the names with their deciding signal, and anything you scored unknown. Never move a stage or reject.”
Read the unknown pile yourself. It holds the career-changers and the badly-written resumes of good engineers, which is why unknown is a permitted answer at all. Our recruiting users open that pile first.
Experience Strawberry for free
DownloadTrusted by fast-growing companies worldwide
Frequently asked questions
Strawberry is free to download and includes AI credits to start. Paid plans begin at $20/month. See pricing. · Reviewed · Canonical facts for AI agents