A flight simulator for analyst work
Practise the job, not the interview.
Spend a week inside a company that doesn't exist. Real extracts with real mess in them, a stakeholder who won't define her terms, a manager who reviews what you send, and a deadline.
You are not graded on whether the query ran. You are graded on whether you noticed the number was answering the wrong question.
Free while in testing · Runs in your browser · No account, no card, nothing to install
The gap is context, not skill
Why practising SQL puzzles stopped being enough.
AI writes the query now. That part is settled. But the jobs that used to teach judgment — the grunt work, the first six months of being handed small things and slowly learning what the numbers meant — are exactly the jobs that got automated. Companies still expect judgment on day one. There is nowhere left to acquire it.
So interview prep teaches the part that has been automated. You can grind a hundred SQL problems on clean, sanitised tables and still freeze the first time a VP asks for "the churn number" and every reasonable definition gives a different answer.
Context Gap gives you somewhere to practise it. Not a course, not a quiz, and not a dataset with an answer key at the back — a week of work inside a company with a history, an org chart, and data that carries the scars of both.
How a session works
You get a link. Everything after that happens in one browser tab, and it saves as you go — close it, come back tomorrow on a different machine, carry on.
You arrive three weeks in
You're an analyst at Kestrel Software, a mid-market B2B SaaS company with ~1,600 customers and a billing migration in its recent past. Your manager sends you the first real assignment. Nobody explains the company to you, because nobody would.
You work, and you ask
Download the extracts or query them right in the page — there's a SQL workbench built in, so nothing to install. Message your manager and your stakeholder like colleagues. They know things they won't volunteer. Asking the right person the right question is half the assignment.
You ship, and get read
You submit real deliverables — a memo, a dashboard spec — and they're assessed against what was actually in the data. At the end you get a debrief: what you caught, what you missed, and what it would have cost.
Two seats
Same company, same week, different chair. Each runs at three levels — the level changes how much is documented for you, not how hard the data is.
Data analyst
Messy extracts, a stakeholder who won't define her terms, and a deadline. You own a number that several people already believe they know.
Mid — fewer answers are documented
Senior — table names, and nothing else
Analytics engineer
Inherit a departed engineer's half-built dbt project. Find what is wrong with it, then write the tests that would have caught it.
Mid — a trimmed README; what is unwired is yours to spot
Senior — no README; the project is what is in the files
What it actually feels like
An illustration of the texture, not a transcript from a real session.
Messages
Nobody in the simulation will tell you that you have asked a good question. The debrief will.
What you're judged on
The data has things wrong with it — the ordinary kind of wrong that every real company's data has, put there on purpose and documented in advance so the assessment isn't a matter of opinion. Some of them change the answer. Some of them look alarming and don't matter. Telling those apart is the whole exercise.
A submission that reports a confident number nobody can defend scores below one that reports a smaller claim with its limits stated. That is not a scoring quirk; it's the thing the job actually rewards, and the thing interviews are bad at testing for.
Questions
Why do I have to ask for access?
Each session is a real conversation with a language model behind it, which costs money to run, so places are handed out one at a time rather than left open to the internet. It also means I can pick the seat and level that fit what you're preparing for.
It's free while in testing. You get a personal link — no account, no password, nothing to install.
How long does it take?
Most of a session is thinking, not typing. Budget two to three hours, and don't feel you have to do it in one sitting — your link resumes exactly where you stopped, on any machine.
Do I need to install anything?
No. There's a SQL workbench in the page itself — a real database over the extracts, running in your browser, so nothing you type there is sent anywhere. If you'd rather use your own tools, the extracts download as plain CSV and there's a short guide to setting DuckDB up locally.
How good does my SQL need to be?
Comfortable with joins and aggregates is plenty. This is not a syntax test — if writing the query is the hard part for you, a technique drill will serve you better than this will. What's being exercised is what you do once the query runs and the number looks odd.
Is my TabsOnJobs login the same thing?
No — they're deliberately separate. Context Gap has no accounts at all; your link is your identity. Having a TabsOnJobs account gives you no access here, and asking for access here creates nothing on TabsOnJobs.
What happens to what I write?
Your messages and submissions are stored against your session so it can resume and be debriefed, and I read them to work out where the simulation is confusing or unfair — that's the point of the testing phase. They aren't sold, shared, or used to train anything. Queries you run in the in-page workbench never leave your browser at all.
Ask and I'll delete your session and everything in it.
Is the company real?
No. Kestrel Software, everyone in it, and every row of its data are invented. The failure modes in the data are real ones, which is the part that transfers.
Request access
Tell me which seat you want and roughly what you're preparing for, and I'll send you a link. Usually within a day.