Ask a language model to 'improve my CV for this job' and you will get something fluent, confident and partly invented. It will add a percentage you never measured, promote you half a level, and hand back a document you cannot defend in an interview.
The fix is not a cleverer model. It is asking for something narrower: find the evidence already on the page, work out which requirements it maps to, and rewrite only the lines where the mapping is unclear. Everything below is that job, split into four prompts.
Get the inputs right first
You need two things in plain text: your CV, and the full posting rather than the two-line summary from the job board. Paste both as text — a PDF pasted from a viewer often arrives with the columns interleaved, and you will spend the session correcting the model's reading of your own document.
Strip anything you would not want stored: home address, phone number, client names under NDA, the referee's details at the bottom. None of it affects the tailoring.
Step one: read the posting, and nothing else
Do not let the model see your CV yet. Ask it to work only on the posting: pull out five to eight core responsibilities, the qualifications split into required and preferred, every named tool or system, and its best guess at the three things that matter most. Ask it to say which terms are genuinely technical and which are generic filler.
The output is a specification. It is also a sanity check on the role itself — if the priorities the model extracts are not the job you thought you were applying for, that is worth knowing before you spend the evening.
Step two: compare, in a table
Now add your CV and ask for a four-column table: the requirement, the exact line of your CV that evidences it, a status, and what to do about it. Constrain the statuses to three — clearly demonstrated, present but unclear, not demonstrated — and tell it to be conservative, marking anything it had to infer as unclear rather than clear.
This is the most valuable output of the whole exercise, and you have not rewritten a word yet. The 'present but unclear' rows are your edits. The 'not demonstrated' rows are your honest gaps, and they tell you whether this role is worth applying to at all.
Step three: rewrite only what the table flagged
Ask for revisions to the summary, the skills block and the specific bullets in the unclear rows — not the whole document. Require it to show, for each change, the original line, the revised line, and the fact on your CV that justifies it. That last column is what stops invention: if it cannot cite a source fact, it should not make the claim.
Tell it explicitly to leave employers, dates, titles and the scope of what you did untouched, and to ask you a question rather than estimate anything it does not know. A prompt that ends 'if a number is missing, ask me for it rather than guessing' is the single highest-value sentence in the whole workflow.
Step four: audit what came back
Start a clean pass and ask it to act as a sceptical reviewer of the revised CV against the posting. It should flag any claim not supported by the original, any metric that appeared from nowhere, any place where your scope or seniority quietly grew, and any keyword dropped in without evidence behind it.
Then read it yourself, out loud, and check three things: every sentence is true, no two bullets have the same shape, and every claim is one you could talk about for two minutes under questioning.
What not to ask for
Some requests are guaranteed to produce a document that will embarrass you later.
- 'Rewrite this as the perfect candidate for the role.' You are not applying as a fictional person.
- 'Add impressive metrics.' It will, and you will be asked about them.
- 'Make sure this passes the ATS' or 'give me a match score'. A chat model is guessing at both.
- 'Fill the gaps in my experience.' A gap the model fills is a lie you have to maintain through three rounds.
- 'Rewrite the whole CV.' You lose the parts that were already working and cannot tell what changed.
The mistakes this workflow is designed around
Two failures account for most bad tailoring. The first is keywords without evidence: pasting the posting's vocabulary into a skills list so the terms are present, which reads as padding to a recruiter and collapses on the first question. The second is uniformity — every bullet starting with 'spearheaded' and ending in a percentage, which is instantly recognisable as machine-written.
Vary the sentence shapes. Leave a bullet without a number if you do not have one. A CV that reads like a person wrote it is doing something the model cannot fake for you.
Or skip the prompt engineering
This is exactly the sequence Apply Engine runs on every application: it reads the posting, maps it against your profile, rewrites only what needs to change, and shows you each version before anything sends — with the source facts still attached to the claims. Nothing goes out without your approval, and every version is stored against the application so you always know what a recruiter is reading.
The prompts above are worth knowing either way. They are the difference between a tool that helps you describe your experience and one that invents a better candidate than you.
