The CV Prompt That Screens You Out First
88% of employers say qualified high-skills candidates get vetted out for not matching the exact criteria in the posting. So don't ask AI to improve your CV — make it screen you out. The prompt, and the three decisions that keep it honest.
- #career
- #resume
- #prompts
- #job-hunting
- #international
Harvard Business School and Accenture surveyed more than 2,250 executives across the US, UK and Germany. 88% of them said qualified high-skills candidates were vetted out of their own process because those candidates did not match the exact criteria set by the job description — 94% for middle-skills roles. Same study: 48% of employers filter middle-skills candidates purely on an employment gap longer than six months, and the business leaders interviewed estimated that filtering disqualified more than half of their high-skills applicants.
Read that as a screener, not as a candidate. The criteria doing the cutting are not some secret standard of quality. They are sitting, in writing, in the posting you are about to apply to. And almost nobody makes the model read them in that order.
The gap the generic advice never closes
When you ask a model to improve your CV, you hand it your document and it optimises for the thing you brought: the best possible version of you. Every bullet gets a stronger verb. Nothing gets cut.
When I open a stack of applications, I am not doing that. The first pass is not reading — it is matching. I have two or three things in my head that the role genuinely cannot teach on the job, and I am scanning for evidence of those. Everything else on the page is noise until that evidence shows up. A polished CV with the proof buried in the fourth role loses to a plain one that puts it in the first two lines.
So the useful artifact is not a rewriter. It is a prompt that puts the model in the chair of the person deciding whether to stop reading.
The prompt
Paste this into Claude (or any chat model), then paste the posting and the CV underneath it.
You are the hiring manager who owns this headcount. Not a recruiter, not a
coach. You have forty other applications open and you are doing the first
pass — deciding who gets a phone screen. You are not here to help anyone.
Work in this order. Do not skip ahead.
STEP 1 — Build the rubric from the posting alone, before you look at the CV.
List the five requirements this posting is actually screening on, in the
posting's own words. Mark each MUST or NICE, and name the signal that told
you which: repetition, position on the page, or the level in the title.
STEP 2 — Screen the CV against that rubric. For each requirement, quote the
exact sentence from the CV that proves it. If no sentence proves it, write
MISSING. Do not infer it from a job title, a company name, or an adjacent
skill. Then state where the evidence sits: which role, and whether that role
is one of the candidate's two most recent.
STEP 3 — Verdict: ADVANCE, BORDERLINE or PASS. Name the single line that
decided it. A PASS must be caused by a missing MUST, never by weak writing.
STEP 4 — List the three sentences on this CV a screener will not read at all
in the first pass, and what each one costs in space.
STEP 5 — Only now, rewrite the three bullets whose evidence is real but
buried. Use only facts already on the page. Invent no numbers, no tools, no
scope. Where a bullet needs a number the CV does not contain, leave [?] and
state what would have to be measured to fill it.
No compliments. No summary of strengths. No encouragement. If you cannot
find a reason to reject, say what evidence would change your mind. Stay
under 500 words before STEP 5.
POSTING:
<paste the full posting>
CANDIDATE CV:
<paste the CV as plain text>
The three decisions inside it
1. The posting goes in first, and becomes the rubric. This is the whole trick, and it is an ordering trick. If the model sees the CV first, it reads the CV and then rationalises the posting against it — suddenly everything looks like a match. Forcing it to extract the criteria before it has seen a single bullet reproduces the real sequence: the criteria existed before you did. Given what HBS found, the exact wording of the posting is not decoration. It is the filter.
2. Ask for a decision with quoted evidence, not a rewrite. "Quote the exact sentence that proves it, or write MISSING" is the anti-hallucination mechanic. A model that has to quote cannot invent — and the MISSING lines are the real output. That is the list of things a screener will fail to find, which is the only list that changes what you do next. The rewrite comes last on purpose, capped at three bullets on purpose, because rewriting is where the model stops screening and starts selling.
3. Take yourself out of the prompt. Anthropic's own research is blunt about why this matters: models finetuned on human feedback learn sycophancy, and "when a response matches a user's views, it is more likely to be preferred" — with humans and preference models sometimes preferring "convincingly-written sycophantic responses over correct ones". Role prompting helps: Anthropic's guidance is that setting a role "focuses Claude's behavior and tone for your use case", and "even a single sentence makes a difference". But a role does not delete the incentive. Removing yourself from the frame does. Write "CANDIDATE CV", never "my CV". Say "the candidate", never "me". The moment the model knows it is talking to the person on the page, the pressure to be nice comes straight back. The length cap earns its place too — Anthropic notes that Opus 5 runs longer than earlier models by default and tells you to "prompt explicitly for conciseness".
My rule of thumb
If the output could be pasted onto somebody else's CV without changing a word, the model was writing marketing copy, not screening. Screening produces a rejection reason with a quote attached to it. No quote, no screen — run it again.
One more thing about the international angle. If you are applying from Brazil to a US or EU company, the criteria you are most likely to miss are not technical. They are scope and recency: what you owned rather than contributed to, at what size, and whether it happened in your last two roles. Those are the sentences a screener reads. When the posting names a level, add scope and recency to STEP 1 as MUSTs so the model has to check them out loud.
Do this today
Take one posting you actually want and one CV you have already sent to somebody. Run steps 1 to 4 only — stop before the rewrite. Read the MISSING list. If a MUST is missing because the evidence exists but sits in your third role, that is a fifteen-minute fix. If it is missing because you have never done the thing, you just saved yourself an application, and that is the better outcome of the two.
Sources
Keep reading
More guides like this one.
The Mock Interview Prompt That Refuses to Help You
The structured interview is the strongest single predictor of job performance, and also the one with the widest spread, because it can be built well or badly. A chat model builds you the bad one. Here is the prompt that installs the structure instead.
Your AI Fluency Collapses Exactly Where the Code Ships
Anthropic classified 9,830 Claude conversations. The ones that produce code get briefed better and audited worse — fact-checking falls 3.7pp. The index publishes no personal score, so here is how to run one on yourself.
Laid Off? These 30 Days Fill Your Calendar, Not Your Résumé
A remote US or EU hiring loop almost never closes inside a month, so a 30-day plan that promises an offer is lying to you. Here is the week-by-week version that ends in a full pipeline instead — and where the free AI credentials honestly belong in it.