Interview prep from the job ad
Mark Dickie · October 2026
How to work out what to study for one specific role, and which tools turn the ad into practice questions
The short version
Most of what you'll be asked is written in the ad, once you read it as a list of interview topics rather than a description of the job. Copy out every technology and every kind of problem it names, separate what it requires from what it would like, and match each item to the round that will test it. Then ask the recruiter on the first call what the technical rounds involve, because the ad rarely says.
That method needs a text editor and nothing else. Some tools will do the first half for you: paste the ad and they hand back practice questions. They differ in one way that matters more than price, which is whether a person or a model wrote the questions.
I run Tarmac, which builds practice quizzes from a pasted job ad. Weigh what follows with that in mind.
What should I study for an interview at a specific company?
Two sources are worth more than the rest put together: the ad for the role, and whatever the company publishes about how it hires. The ad tells you the subject matter. The company's own page, if it has one, tells you the format: how many rounds, how long each is, and who you'll meet.
Candidate reports on Glassdoor, Blind or Reddit come third. They're good for format and timing. They're poor for the questions themselves, because a reported question is one person's memory of one interviewer on one day.
Skip lists of "the 50 most common interview questions" for the role. They aren't wrong, but they're the same list for every company, and the point of this exercise is the part that isn't.
How do I turn a job ad into a study list?
- Copy out every named technology, exactly as the ad writes it. Languages, frameworks, databases, cloud services, tools.
- Copy out every kind of problem the ad names: "performance", "scale", "reliability", "data pipelines", "security". These become the system design and debugging questions, and they're the part people skip.
- Split the two lists into required and nice to have. Ads usually separate them under headings. When they don't, anything in the first paragraph or repeated twice counts as required.
- Put each item next to the round that will test it: a coding round, a system design round, questions on the stack, or a behavioural round. The bullets about communication and ownership belong to the behavioural round. Write one true story from your own work for each.
- Mark every item by what you've done with it: used it at work, used it on your own, or never touched it. Study the required items you've never touched first, then the required items you've only used on your own.
- Divide the days you have left by the number of items in your first two groups. If the answer is under a day each, drop the nice-to-haves.
A worked example: a backend engineer at GitLab
GitLab publishes its job descriptions in its public handbook, so its backend engineer page is a real ad you can read alongside this.
The shared requirements for every backend role run to 17 bullets. Only a few name something you can study. The first is "Significant professional experience with Ruby on Rails or language required by the specialty". The page then says why Ruby matters even when a specialty uses another language: "Because the majority of our codebase is written in Ruby, even in specialized roles, a knowledge of Ruby and Ruby on Rails or a desire to learn and contribute is necessary."
One bullet names a kind of problem: "Experience with performance and optimization problems and a demonstrated ability to both diagnose and prevent these problems". Another names a practice: "Comfort and familiarity with our code review process". Most of the rest describe how you work with other people.
Here is the study list that falls out of it.
| From the ad | Round it points to | What to study |
|---|---|---|
| Ruby on Rails | Stack questions, coding | Reading and changing unfamiliar Rails code; ActiveRecord queries and callbacks |
| Performance and optimization problems | System design, debugging | Query plans and indexes, N+1 queries, caching, finding a slow request from its logs |
| Code review process | Coding, or a review exercise | Writing review comments on a diff: what you'd block, what you'd let through |
| Communicating complex technical problems; owning a project from concept to production | Behavioural | One story each, with what you decided and what went wrong |
The page also sets out the process: a 30-minute screening call with a technical recruiter, a 90-minute technical interview with a backend engineer, then a 60-minute interview with an engineering manager and another with the director of engineering. It doesn't say what the 90 minutes involve. That's the question to ask in the screening call: is it live coding, a take-home walked through, a review of existing code, or a design discussion. The answer changes how you spend the next week more than anything else on the page.
What a job ad doesn't tell you
An ad says what the team works with. It doesn't say how deep each item will go, or whether there's an algorithm round at all. Some companies still run timed puzzle rounds for roles whose ads never mention algorithms, and some have dropped them for roles that do. The recruiter knows, and asking is normal.
An ad also can't tell you which of its requirements you already meet well enough. That's the one part of this you can only find out by answering questions and checking the answers.
Which tools turn a job ad into practice questions?
These are the tools I could find that take a pasted job ad or job link, checked on each one's own site on 5 October 2026. None of the coding or system design platforms (LeetCode, NeetCode, AlgoExpert, Hello Interview) takes a job ad.
| Tool | Where the questions come from | Runs your code against tests | What it costs |
|---|---|---|---|
| Interview Forge | Written by a model from the ad and your CV | No | Free for up to 50 questions a day |
| InterviewPal Job AI | A model reads the ad and selects questions; the site doesn't say whether they're generated or collected | No | $29 once, for life |
| Duos | A model builds the company, the team and the questions from a job link and your CV, as a spoken mock interview | No | One interview free, then $9.99 for 24 hours or $19.99 a month, renewing |
| Huru | A model writes questions from the pasted ad, alongside a library of 50,000 | No | Free for job seekers |
| Yoodli | Its own question bank or your list, set against the job details you give it | No | Free for 5 role-plays, then from $8 a month billed yearly |
| Final Round AI | Reads the ad into a prep checklist. Sold mainly as an assistant that runs during the real interview | No | Ad reading free; from $25 a month billed yearly, or $150 for one month |
| Tarmac | Picked from a bank of 9,302 questions, each fact-checked and reviewed by a person | Yes, with an account | Free to start; Pro is one payment from $19 |
Most of these are built for spoken and behavioural rounds: you answer out loud and get a score for clarity and structure. If the ad points to a coding round, check that the tool can run your code before you pay for it.
Are AI-written interview questions reliable?
A model can write questions for any stack, including ones no fixed bank covers. That's a real advantage when your ad names something unusual.
The cost is that you can't easily tell when a question or its model answer is wrong, because you're the one learning the material. A study presented at CHI 2024 checked ChatGPT's answers to 517 Stack Overflow programming questions and found that 52% contained incorrect information. In the same study, programmers shown those answers missed the mistakes 39% of the time. That was the 2023 version of ChatGPT. Today's models are better, but I haven't seen the same check repeated on them, so treat a generated answer as a lead to verify.
Tarmac makes the other trade. Every question comes from a fixed bank of 9,302, across 76 areas, and each one was fact-checked and reviewed by a person before it went in. A model reads the ad once to pull out the skills it asks for. The questions are then picked from the bank, and no model writes one. The limit is coverage: if your ad's main stack isn't one of those areas, Tarmac has less to give you and a generator will cover it.
How Tarmac does it
Paste the ad at gettarmac.com/practice. Before the first question you see the topics the ad asks for and how many questions Tarmac holds on each. Then you get a 10-question quiz built from them, and the same ad builds the same quiz every time.
The first 3 answers need no account, and you can do this for 1 pasted ad a day. Without an account the quiz leaves out coding questions, because those run your code against tests and need a sign-in. A free account adds 2 job-ad quizzes every 30 days.
Pro lifts the caps: $19 for a month, $49 for 3 months or $129 for a year. Each is one payment that stops on its own. Tarmac reads job ads from the United States, the United Kingdom and Australia, so its figures on what ads ask for cover those three.
Sources
- GitLab, Backend Engineering job family, GitLab Handbook, read 5 October 2026.
- Samia Kabir, David N. Udo-Imeh, Bonan Kou and Tianyi Zhang, Is Stack Overflow Obsolete? An Empirical Study of the Characteristics of ChatGPT Answers to Stack Overflow Questions, CHI 2024.
- Interview Forge, home page; InterviewPal, Job AI and pricing; Duos, home page; Huru, home page; Yoodli, pricing; Final Round AI, subscription page. All read 5 October 2026.