Which ChatGPT prompts actually check a CV against Swiss standards?

Why the Swiss standard has to be inside the prompt

A language model has read far more American application documents than Swiss ones. Given no standard it therefore checks against whichever norm dominates its material: one page, no photo, no residence-permit line, no language levels. Those three points are customary to expected in Switzerland — and a model that reports your photo as a defect has not reviewed you, it has translated you into the wrong market.

So the standard goes in the prompt, not in a follow-up message. Five facts are enough: length two to three pages, photo customary, permit status visible, every language with a level from A1 to C2 or native, and a workload in percent where it matters. That is the yardstick a Swiss reader actually applies.

Languages are the point most people underrate, and our own numbers show why: in nursing adverts 35.2 percent name no language requirement at all, in education 38.6 percent. Unstated does not mean unexpected — it means the expectation is taken as read and tested in the interview instead. Anyone who does not write their own level down has handed that assumption to somebody else.

Prompt 1 — the Swiss baseline check

You are an experienced Swiss hiring manager with fifteen years of recruiting for the Swiss labour market. I am pasting my full CV as plain text below. Review it against the Swiss dossier convention, not against a US resume: two to three pages is normal here, a professional headshot is customary and is not a mistake, work-permit status (B, C, L, EU/EFTA or Swiss citizenship) belongs visibly in the header or under availability, every language needs a level on the European framework from A1 to C2 or the word native, and a workload in percent is worth stating whenever I am not looking for a hundred percent role. Give me a table with exactly these columns: element, what you found, a grade of present / partial / missing, and one sentence of concrete correction. Add nothing that is not in the text, and flag every point where you had to guess. My CV: [paste the text here]

Three elements do the work here. The role sets the standard: a hiring manager reads differently from a career coach, because a hiring manager is looking for reasons to put a dossier down. The list of Swiss conventions stops the model from checking against a US resume. And the three-grade table forces a decision per point — without it the model returns a balanced paragraph on every element, in which nothing is wrong and nothing is usable.

The do-not-invent sentence is not decoration. A model that finds a gap likes to fill it — with a plausible year, a plausible qualification, a plausible team size. In an application dossier that has stopped being a question of wording.

Prompt 2 — matching the advert

You are a Swiss recruiter reviewing one application against exactly one job advert. I am giving you two texts: A the advert, B my CV. Work in three steps. First: extract the twenty most important requirement terms from the advert and rank them by weight, putting everything listed as a must or a requirement ahead of the nice-to-haves. Second: for each term, say whether my CV contains it verbatim, in substance, or not at all, and quote the passage for verbatim and in-substance matches. Third: name the five gaps most likely to drop me in the first round, and for each say which section of my CV the fix belongs in. Do not rewrite anything, only propose. If you need Swiss labour-market context, fetch https://swissjobs.app/llms.txt; the prompt works without it too. A: [paste advert] B: [paste CV]

This is the only one of the five that decides an invitation, and it differs from the others by one ingredient: the text of the role you are applying for. The model has never seen that advert. Half the open roles in our index were under three weeks old at the time of the analysis — in engineering the median advert was fifteen days old, in construction forty-one. Nothing that young was in any training material.

The three steps are deliberately separated and ordered. Terms from the advert first, then the match, then the ranking. Drop the steps and the model jumps straight to recommendations and justifies them with words the advert never used. The instruction to quote the passage for every hit is the built-in control: a match you cannot quote was not a match.

The instruction not to rewrite has a practical reason. The moment a model reformulates your text you have two jobs instead of one — close the gaps, and check what the rewrite quietly lost. Accepting suggestions is faster than proofreading a new version of your own CV.

Prompt 3 — the format check for applicant tracking systems

Review the following CV for machine readability only — not for content and not for qualification. For each of these nine points give me a grade of present / partial / missing, one sentence on what you found, and a concrete fix: contact details as real text rather than only inside a header graphic; conventional section headings; single-column layout with no sidebar; the exact job title I am applying for present verbatim in the document; one consistent date format throughout; a readable text layer rather than a scanned image; length within the Swiss norm of roughly two to three pages; coverage of the technical vocabulary of my field; photo. Treat letter-spaced headings — headings written with spaces between the individual letters — as partial, because they break word tokenisation. Finish with one sentence naming the single point that costs me the most. CV: [paste here]

This prompt explicitly does not check qualification, which is why the word only appears in the first sentence. Mix form and content in one request and the model's attention drifts to the content, because there is more to say about it — the format faults then turn up as a subordinate clause, or not at all.

The nine points are not arbitrary: it is the same list our own check works through, and each names a way a document loses information on the way in. Contact details inside a header graphic do not exist to a parser. A two-column layout is read line by line and the columns interleave. Letter-spaced headings fall apart into single characters.

One limit no prompt removes: the model sees your text, not your layout. It infers a second column only from the order in which the words reach it. Hence the closing question about the single most expensive point — it forces a ranking instead of nine equally weighted remarks.

Prompt 4 — turning duties into results

You are a strict editor of application documents. Take the following duty descriptions from my CV and rewrite each one as a measurable result. The rule: every line starts with a verb, names a number or a defensible order of magnitude, and ends with the effect rather than the task. Where I have not given you a number, ask me for it instead of inventing one — put up to five targeted questions to me and wait for my answers before you write. Keep the finished lines plain; delete superlatives and marketing language. Then draw three to four lines from the same material for a short profile that says in one sentence what separates me from the other people applying for this same role. My duty descriptions: [paste here]

The most effective part of this prompt is the instruction to ask rather than invent. Without it every model produces numbers that sound right and are not, because the task demands a number. With it you get five questions back, ten minutes of thinking, and a result that is actually yours.

The second half — the short profile — matters more in Switzerland than its length suggests. On a well-cut role somebody reads many dossiers in a row, and the first three lines decide whether yours is read carefully. What stands there has to be true of you and not of the others.

Where this prompt pays most is visible in our seniority figures: in marketing 45.1 percent of adverts carry a seniority signal in the title, in IT 28.7 percent, in commercial and administrative roles under two percent. Where the level is advertised it is also examined — and it is examined on evidenced results, not on years.

Prompt 5 — register, tone and local usage

Review the language of the following application text for the Swiss market. Check four things and give me the affected passages with a proposed correction for each. First, register: Swiss application writing is more restrained than the Anglo-American norm, so mark every sentence that oversells and propose a flatter version. Second, spelling and vocabulary: use the Swiss variant where German, French or Italian differ from the neighbouring country's usage, and name the local term for qualifications and workload. Third, filler: mark every sentence that could appear in any application at all and propose a version that is only true of me. Fourth, model voice: mark phrasing that reads as written by a language model and replace it with something plainer. Change no facts. Text: [paste here]

This prompt solves something the other four cannot see: a dossier can be factually right and still sound like it came from somewhere else. The gap that catches most English-speaking applicants is register. A line that reads as confident in a London or New York application — a superlative, a claim of transformation, an achievement stated in the first person plural — reads as overselling in a Swiss one, and restraint is not modesty here, it is credibility.

The fourth point, model voice, is new and getting more important quickly. Draft an application in a chat and you inherit its sentence rhythm — and by now that is noticed even by people who have never asked themselves why a text sounds the way it does. Running the same check a second time with the instruction to delete anything a model might have written is the cheapest quality control in the whole sequence.

If you are applying in French or Italian Switzerland, run this prompt in that language rather than in English. It is worth it twice over: in hospitality 20.7 percent of adverts name French, in commercial roles 18.9 percent. A dossier submitted in the second national language will also be read in it.

What a chat cannot do in this review

Three limits are fixed and no prompt lifts them. First, a chat does not see your layout: you paste text, so it judges text. Whether your second page is half empty or your type is too small, it cannot know. Second, it knows the advert only if you paste it in — and the roles that matter are usually younger than its training material. Third, it has no judgement on whether a formulation will hold up in an interview; only somebody who knows the industry has that.

So the sensible order is machine first for the mechanical part, a person last for the rest. In Switzerland that human feedback is also cheap to come by — the regional employment centres include it in what they offer.

What not to paste into a chat

A CV is a dense personal document: address, date of birth, permit number, former employers, sometimes a photo. Before pasting, it is worth checking the provider's own settings for whether your input is used for training, and worth a second look at what you actually need: every one of the five checks works on the content without your home address, your date of birth or any identity-document numbers.

Former employers and dates do belong in, or nobody is reviewing anything meaningful. Contact details of referees do not — those are other people's data, and nobody delegated that disclosure to you.

The order in which the five are worth running

Start with prompt three, the format check. It is the fastest, and the only one whose findings you can fix without thinking: one column, one date format, one heading. Then prompt four, because results instead of duties genuinely change the text and everything else builds on it. Then prompt one for the Swiss convention, and prompt five for the language.

Prompt two runs last, and it runs again every time. It is the only one tied to a specific advert — the other four you do once, this one you do per application. That is also exactly why we built it into our own check: it is the one people skip when they are doing this by hand.