Optimising a CV does not mean reinventing it. It means arranging and phrasing the same material so that what you can do is visible in five seconds. Language models are well suited to that, because the material already exists and little has to be invented.
This article walks through the process, where AI genuinely helps, and where you have to read over it yourself. The fundamentals of the document are in the German CV guide.
Optimising is not reinventing
The most common mistake is telling the model to "make my CV better". What comes back is a different CV: with words you never use, roles that sound slightly larger, and task descriptions you would not recognise in an interview.
It becomes useful when you set small, clearly bounded tasks. Not "improve this", but "rewrite these five task lines as results without adding anything". What you put in determines what comes out, and what you leave out determines what gets invented.
Finding the weak spots
Before rewriting, run a diagnosis. Four patterns sit in almost every CV that grew over time, and a model finds them reliably when asked.
Passive constructions. "A new till system was introduced" does not say what you did. In the active voice your part is in the sentence.
Task lists instead of results. "Responsible for travel expense accounting" describes a job description, not an achievement. What came out of it?
Missing magnitudes. How many sites, how many cases, how large the team. Those numbers are almost never there and make the biggest difference.
Gap logic. Jumps in time, unclear transitions, a move with no visible reason. A model spots them immediately, because it only reads the dates.
A usable prompt for that:
Read this CV and list: every passive construction, every task description without a result, every entry without a size, and every unclear date range. Do not rewrite anything.
Working keywords in from the ad
Applicant tracking systems search your CV for terms. If the ad says "Reisekostenabrechnung" and you wrote "processing of business trip expenses", the search does not find you. How those systems work is in the article on the ATS-proof resume.
The procedure is mechanical:
- Put the ad and the CV side by side and have the terms marked that appear in the ad and are missing from yours.
- Check each one: does it apply to you? If not, it stays out.
- Build the applicable ones in, in exactly the ad's spelling, where they belong in substance.
An example. Before:
Handling of business trip expenses and checking of receipts
After:
Reisekostenabrechnung for around 60 employees, including receipt checks and clarification with accounting, roughly 120 cases a month
The same activity, now with the ad's term and a magnitude. Note that German system and process names stay in German even in an English CV: nobody searches for the translation.
Quantifying results
The most effective step and the most uncomfortable, because the numbers are rarely written down anywhere. A model helps here not by supplying numbers but by asking the right questions.
A useful prompt:
Ask me three questions about each of these stations that could help me find a measurable figure. Do not invent any numbers.
Typical figures almost everyone has and nobody records: number of sites looked after, team size, cases per day or month, number of systems, budget scope, number of colleagues trained, project duration.
Where you genuinely have no number, use a checkable description. Invented percentages are the worst trade available: they read well until somebody asks in the interview.
Polishing
Three short passes at the end.
Tone. Consistently factual, no superlatives, no adjective chains. A CV is a working document, not advertising copy.
Length. One page for beginners, at most two with experience. Cutting means deleting, not setting smaller type.
Consistency. One date format, one spelling per employer, one tense per block type. Models are good at this because it is diligence work.
What AI should not touch on a CV
Three areas are better left alone, because there is nothing to gain and something to lose.
The structure of the German CV. Reverse chronological, a header with contact details, then work experience, education, skills. Models like to suggest an objective at the top or a competency-based ordering; both are unusual here and read as constructed.
Personal details. Name, address, date of birth and photo do not belong in a stranger's input field, and they need no phrasing anyway. Rewriting work needs activities, date ranges and numbers.
Gaps. A model that sees a jump in time likes to smooth it away by stretching date ranges. That is the point where phrasing help turns into a false statement. Gaps get explained, not disappeared.
A full pass
Here is the work on one station from start to finish. Starting point:
Purchasing administrator, Müller GmbH, 2019 to present Responsible for orders, supplier contact and invoice checking.
The diagnosis finds: no magnitude, no result, no system names. After rewriting with your own details:
Purchasing administrator, Müller GmbH, 03/2019 to present Procurement for three production sites, around 200 orders a month in SAP MM. Supplier selection together with production, invoice checking through to release. Training new colleagues in the team since 2023.
Same role, same tasks, nothing invented. What came in are four details you supplied yourself, plus a system name a tracking system can find.
Limits: what you check yourself
Three things stay with you, and together they take ten minutes.
Every number. Not "sounds plausible" but "is true". Models like to round.
Every title and date range. "Deputy" becomes "head" quickly, and a gap becomes a smoothly running timeline.
The selection. What belongs in the CV at all and what can go is your call. A model does not know the job you are applying for, let alone your priorities.
Do those three checks and the rest is time saved. Jobvin's AI resume is built for exactly this: it works on your stored profile, rewrites tasks as results, and produces a tailored version per application without changing your base version.
Frequently asked questions
Will AI distort my CV?
It can, if you let it. Models happily turn "contributed to" into "was responsible for" and round numbers. Which is why the last step is always checking against reality, line by line.
How often should I optimise?
The base version once, thoroughly, then a small adjustment per application. After the first pass that takes minutes, because only the summary and the ordering change.
Does Jobvin tailor per application?
Yes, that is the point: a version cut to the ad comes out of your master profile without you touching the base version.
What stays manual?
Choosing which stations belong in at all, checking every number and date range, and deciding what to leave out. No model can make those three calls for you.