Skill tests on job boards are supposed to separate the strong candidates from everyone else. A developer who posts as vikvlg just showed how thin that line can be: he let Google's Gemini walk him through hh.ru's professional skill tests in a subject he openly admits he doesn't know, and the attempts counted. hh.ru is one of Russia's biggest job sites, and these aren't throwaway quizzes — they feed into the country's national IT competency system, and passing them gets you an electronic certificate through the Gosuslugi government services portal. Which raises the real question: not "can AI pass a test," but "what are these tests actually measuring in the first place?"
What happened
It started as a copy-paste problem
Viktor, the developer behind the write-up, wasn't setting out to cheat. He wanted a database of the questions. Copying them by hand was off the table: hh.ru watches for tab switching, and if you flick to another window mid-test your result gets voided. Recording the screen and running the video through text recognition felt like homework.
So he built something that reads the page in the background. The setup was a Tampermonkey userscript — a small script that a browser extension runs on a page — which parses the question text and sends it to a local server. Then came the obvious follow-up thought: while this thing is running, why not bolt an AI onto it?
Two functions, then a local model
The first version of the server did exactly two things: save the full HTML page, and save the test text into files. Then Viktor added support for local AI models through LM Studio — a free desktop app that runs models on your own machine instead of in someone else's cloud — so he could fire the test questions at different models and compare answers.
Here's the part that matters. Viktor's background is Delphi, C++ and Java/Kotlin. He says plainly that he doesn't really know web development, JavaScript, Python or AI. He got the whole thing working by chatting with the browser version of Gemini in a loop: follow the suggestion, run the script, paste back the error message, get the next fix. His own edits, he says, were minimal. He also deliberately left the ready-to-run automation code out of the article, framing the experiment as research.
The prompt did the heavy lifting
The system prompt — the standing instruction you give a model before it sees any question — carried most of the weight. Four rules did the work:
• Name the subject. "Testing in Git: one question and several answers follow." A little context goes a long way.
• Don't re-check more than once. Models that audit themselves in a loop burn minutes you don't have.
• No clarifying questions. The thing has to run unattended.
• Answer with the chosen option's text and number only, no explanation. Keep the output clean.
Cheap, sensible, and it works on questions the model has no business knowing the answers to. That's the uncomfortable bit.
What it means for you
If you're job hunting
A certificate with your name on it used to mean a human sat down and proved something. Now it can mean a human typed a good prompt. Assume hiring managers know this. Your best move is to pair the certificate with evidence that survives a conversation — a portfolio piece, a repo, a live walkthrough of something you actually built.
If you're screening candidates
Skill tests are now a spam filter, not proof of skill. They're still handy for cutting a giant pile of resumes down to a shortlist, but treat the survivors as "worth a call," not "verified." A short live task tells you more than any badge.
If you run a small business
Same rule applies when you buy an assessment tool. If a vendor's demo is a multiple-choice test, ask what stops a candidate from pointing an AI at it. Anything that runs in a browser and reads the screen can be read by something else.
If you're studying
Flip the trick around. The exact prompt discipline above — state the topic, demand one answer, forbid the essay — turns a chatbot into a much better tutor. You get a straight answer and then you go verify it, instead of wading through five paragraphs of hedging.
If you make things
The build itself is the takeaway, not the test-passing. A browser script plus a local model is a two-piece pattern you can point at anything repetitive: sorting a folder of screenshots, tagging support tickets, drafting quiz questions of your own. Local models are free to run and your data never leaves the laptop.
If you're chasing income
This is what "AI skills" actually looks like right now — not writing prompts, but gluing a script to a model and shipping something that saves time every day. That's a freelance service you can sell to a local business this month.
How to try it right now
You don't need to automate a test to learn the pattern. Start with the free tier and work up.
1. Free, zero setup: open Google Gemini in your browser and paste in a question from anything you're studying. Add the four rules above, word for word. Compare its answer to the one you'd have given.
2. Add browsing muscle (free): install Tampermonkey in your browser. It's a free extension that runs your own little scripts on pages you choose — that's the piece that read the test page.
3. Write the thinnest useful script: have it grab the heading and body text of a page you're allowed to read and log it to the console. A handful of lines is plenty for day one.
4. Run a local model (free): install LM Studio, download a model, load it, and start its local server. Now you've got an AI answering on your own hardware.
5. Make them talk: a short Python script that takes the text your userscript posted and sends it to LM Studio's local address is the whole bridge.
6. Check the output like a skeptic. The author deliberately didn't publish working code, and you shouldn't skip this step either — log every answer, spot-check a few by hand, and keep a human on the final click.
If you'd rather start with something that needs no wiring at all, MyKreaTool collects free AI tools you can test in a browser tab today.
Upsides and what changes
The barrier is gone. What used to require a team of engineers now takes one person, a free extension, a free app and a weekend. Viktor reckons a beginner could replicate it and an experienced developer could do it a hundred times better.
For employers, that means the certificate arms race is over. The next round is live assessment, recorded interviews and paid trial tasks — slower and more expensive, but much harder to fake.
For everyone else, the upside is the same pipeline pointed at dull work: pull text out of something, classify it, and let a model draft the response while you approve it.
Limitations
Honest version: this is one person's experiment, not a study. No accuracy numbers were published, and the code was deliberately withheld, so nobody can reproduce the exact results. Local models are slower than cloud ones and take real setup time; the same script that beat the test can get you banned from the platform, which is a fine way to lose a job offer you actually wanted. And the deeper hole is this — passing a test proves a model can pattern-match, not that you can do the job. Wave a certificate at a role you can't perform and the interview will find you out. The anti-cheat arms race cuts both ways, too: sites will get better at sniffing out automation, and plenty of legitimate uses of the same technique will get caught in the net alongside the cheats.
Conclusion
A free chatbot, a free browser extension and a free local model were enough to get through a professional test that's meant to certify competence. That's not really a story about cheating — it's a story about what a certificate is worth once the cost of faking one drops to zero.
One thing to do today: take a test question from your own field, hand it to Gemini with the four rules from this article, and see how far it gets. You'll learn more about where AI actually stands — and where your own edge is — from a session of that than from a month of headlines.
Comments 0