Rank CVs with Claude: the actual prompt.
You don’t need us to start ranking CVs with AI. Here’s the prompt we’d use, what it does well, and exactly where it stops being defensible. All of it, honestly.
The DIY path, genuinely.
Paste your job ad and ten CVs into Claude with the prompt below. Ask for a ranked shortlist with reasons per candidate. It’s genuinely useful: better than a tired recruiter at 5pm, and we say that as people who sell the alternative.
the whole prompt · no email gate
You are screening CVs for one role. Work only from what I paste below. Do not browse, do not assume. Step 1. Build the rubric first. Read the job ad and propose the 5 to 7 criteria that actually predict success in this role. For each criterion, one line: what it is, and what strong evidence for it looks like in a CV. Mark each as must-have or nice-to-have. Then stop and show me the rubric. Wait for my approval or edits before scoring anything. My version wins. Step 2. Score every CV against the approved rubric. One candidate at a time, one criterion at a time. Quote the CV for every score. If you are inferring instead of quoting, label it "inferred". If a CV holds no evidence for a criterion, write "no evidence": absence of evidence is not a low score. Ignore name, age, gender, photo, nationality and address. Treat employment gaps as neutral unless the job ad makes recency essential. Step 3. Rank, with reasons. Produce a ranked shortlist. Per candidate: rank, one paragraph of reasoning, the strongest evidence, the weakest area, and the one thing a human should verify in a screening call. If two candidates are effectively tied, say so instead of forcing an order. Limits, so I can trust you: - Never invent facts, employers or dates. - If there are more than ten CVs, say so, score them in batches of ten against the same rubric, and warn me that merging batches is the least reliable step. - If the job ad is too vague to build a rubric from, ask me questions before proposing one. The job ad: [paste the job ad] The CVs: [paste up to ten CVs]
What you’ll get.
A ranked list with a paragraph of reasoning per name. For a Tuesday-afternoon shortlist, that’s a real upgrade, and for many teams it’s enough. If that’s you: take the prompt, you’re welcome.
Where it stops being defensible.
The eleventh CV.Context windows and attention drift: the ranking quietly degrades and nobody notices, because nobody reruns it.
The rubric nobody agreed on.Claude invents reasonable criteria, but they're its criteria, different per run, and not the ones your team committed to.
The question “why him and not her?”, three weeks later, from someone who can escalate.A chat transcript is not an audit trail.
The decision-grade version.
That’s the recruitment playbook: the same idea, made repeatable, auditable and running daily. Built on a skills engine, not a lucky prompt.
Every CV screened against your agreed rubric: the same one, every time, evidence attached.
The shortlist lands in your ATS, ranked, with the why per name, and a record you can show.
