Which prompts automate a job search in Switzerland with AI agents?

What an agent may take over — and what it may not

The line runs between finding and deciding. Searching, sorting, removing duplicates, holding requirements against your own CV, building a weekly overview: those are mechanical jobs, and an agent does them faster and more evenly than you do on a Sunday evening.

What the agent does not take over: sending the application, passing its text off as your own voice, and filing your job-search efforts. The first costs you your reputation, the second your credibility, the third possibly money. All five prompts below are therefore built to end in a draft or a list that you read, change and send yourself.

Prompt 1: the weekly search brief

The first prompt matters most, because it decides what the agent looks at at all. Three things make it good: a profile written in numbers rather than adjectives, a cut-off date, and a source the agent can build valid search addresses from instead of guessing them.

You are my research assistant for the Swiss job market. First read https://swissjobs.app/llms.txt and use the URL patterns described there to build search pages per role and city. My profile: role [job title], workload [80-100%, called Pensum in Switzerland], cantons [Zurich, Zug, Aargau], languages [English native, German B1, French A2], permit status [Swiss citizen / B permit / C permit / needs sponsorship], earliest start [date]. Find open adverts published since [date of my last run]. Return a table with these columns: title, employer, location, publication date, link, languages required, workload, one sentence on why it fits. Rules: only adverts you actually opened; never invent a link; if a field is not stated in the advert write 'not stated' instead of guessing; sort by publication date, newest first.

Why the file comes first: https://swissjobs.app/llms.txt sets out the URL patterns per language — swissjobs.app/jobs/{role}/{city} in English, /stellen/ in German, /emplois/ in French, /lavoro/ in Italian — together with the role and city names that actually exist. Without a template like that, a model invents addresses that lead nowhere and you only find out when you click. The instruction to write 'not stated' rather than guess is the second safety pin: it turns a hallucination into a visible gap you can check.

The cut-off date is not a detail. A good half of the open adverts in our index have been online for more than a month, so without a date filter the agent reads the same stock again every week and hands you a list that is roughly three-quarters old acquaintances.

Prompt 2: matching adverts against your own profile

The second prompt answers the question that follows the search: which of these do I actually apply to. The point is that the score is broken into separate criteria. A single number — 'fit: eight out of ten' — cannot be checked; four numbers show you whether it fails on the language, the workload or the commute.

You are an experienced Swiss hiring manager. Here is my CV: [paste]. Here are [number] job adverts: [paste]. Score each advert from zero to ten on four separate criteria: professional fit, language requirement against my level, workload, and commute from [where I live]. For each advert name the one requirement I clearly do not meet and the one I clearly exceed. Do not invent a requirement that is not in the text: if the advert states no language level, say so. End with three lists: 'apply now', 'apply only with a reworked dossier', and 'do not apply, because'. Give one sentence of reasoning per advert.

The language requirement earns its own column because in Switzerland it decides more rejections than the professional fit does: about two thirds of the open adverts in our index name at least one required language, and the large majority of those state a professional working level rather than conversational.

The clause forbidding invented requirements is not politeness towards the machine, it is the substance. A model that fills in a missing detail will screen you out of roles you are qualified for, silently.

Prompt 3: duplicates and reposted adverts

Anyone harvesting several sources gets the same role several times. In our own index, twenty-seven per cent of open adverts share their title with at least one other advert — and the larger part of that, a good twenty per cent of all adverts, is the same title at the same employer.

Clean up this list of adverts. Treat two entries as the same advert when the employer and the job title match, even if the location or the link differs. Keep the entry with the oldest publication date in each group and add a column 'reposted' saying how many times the same role appears elsewhere in the list and since when. Do NOT treat the same title at different employers as a duplicate. Output the cleaned list, then a short note on which roles appear to have been advertised repeatedly for months. List: [paste].

That is why the rule in the prompt is employer AND title. Deduplicating on the title alone would be wrong — two 'Project Manager' adverts at two different companies are two jobs. Deduplicating on the link alone would be wrong too, because the same role carries different addresses on different portals.

The 'reposted' column is the real prize. A role that keeps reappearing month after month tells you something about turnover or about expectations there — and it is a good reason to ask the employer directly instead of filling in the same form a fourth time.

Prompt 4: tailoring to a single advert

This is where automation stops and preparation starts. The prompt deliberately does not produce a finished letter; it produces a comparison — requirement on the left, evidence or gap on the right. That is the shape that helps you write and that you can verify in two minutes.

You are my application coach for Switzerland. Advert: [paste the advert text]. My CV: [paste]. First, build a two-column comparison: every requirement named in the advert on its own row, and next to it the concrete evidence from my CV or the word 'gap'. Second, for the three strongest matches write one bullet each in the language of the advert, with a measurable result, two lines maximum. Third, name the two gaps I should address in the motivation letter, and how. Hard rule: do not invent a role, a qualification or a result. If the evidence is missing, say so rather than writing something plausible. Respect Swiss conventions: workload as a percentage, languages on the CEFR scale, permit status on its own line.

The hard rule against invented roles is there because generative models are good at exactly that: writing a plausible line that does not survive the interview. In a Swiss process built on Arbeitszeugnisse — written employer references — and on reference calls, that surfaces late, when it is expensive.

The three Swiss conventions in the prompt — workload as a percentage, language levels on the CEFR scale, permit status on its own line — are the difference between a dossier that reads as local and one that reads as an American resume translated.

Prompt 5: the weekly record for your own file

The fifth prompt is bookkeeping, and it is the part almost everyone leaves until the end of the month.

Summarise my applications from the past week. Raw data: [date, employer, role, channel, status, reply]. First give a table with columns date, employer, role, type of application, outcome. Second, three lines of analysis: where I did too little, which channel actually produces replies, and the two follow-ups due this week. Do not rephrase anything into something I did not do. Important: this is my own working record, not a submission. I file my personal job-search efforts myself through Job-Room or with my RAV.

The last sentence of that prompt is the most important on this page. Proof of personal job-search efforts is submitted online through Job-Room, and arbeit.swiss lists that submission explicitly as an eService open only to people who have signed on with the RAV and registered in Job-Room. An agent can prepare the list; you file it, and what counts is what arrives there.

If you keep your applications with us anyway, this prompt is redundant: the application tracker fills the SECO form from the applications you saved and you print it.

How often to run it — what the numbers say

The most common wrong setting is daily. Of the open adverts in our index on the reference day, just under three per cent had been published within the last day, a good eleven per cent within three days, and just under a quarter within a week. The median open advert is about a month old.

That gives a practical cadence: twice a week. A daily run shows you a very thin list and still creates work every time; a run every three or four days catches essentially everything that is genuinely new. Searching daily mostly means searching the same adverts again.

The exception is the reporting-requirement head start below — there it is not your cadence that decides, it is your access.

The blind spot: the first five days

An agent reading public job portals sees part of the market systematically later than other people do. Vacancies that employers register under the Swiss job registration requirement are, in the words of arbeit.swiss, only visible to jobseekers who have signed on with the RAV and registered in Job-Room for the first five days after publication; only afterwards may they be published on other job platforms.

That is not a weakness of the prompts but a property of the market, and it has one simple consequence: if you are registered, look inside Job-Room yourself during those days. No agent, no portal and no tool can substitute for that head start.

What you should never automate

First, sending. An application that goes out without you reading it will eventually carry the wrong company name — the one class of mistake that disqualifies instantly here.

Second, volume. The reflex to use an agent to apply more rather than better is understandable and counterproductive: recruiters in Switzerland recognise the standard letter that would fit a fiduciary office and a care home equally well.

Third, the filing with the RAV. And fourth, your personal data: whatever you paste into a chat window leaves your computer. Home address, date of birth, social-security number and scanned references do not belong in there — a match against an advert needs your experience and your qualifications, nothing more.