Three questions. Anything that didn't serve them was cut.

1. What triggers and what blocks the start of prep in a short window?
2. What does the market actually sell, and what behaviour does it assume?
3. What would a product organised around days remaining have to do differently?

The segment was defined by behaviour, not by age or geography: four to eight weeks to the exam, nothing done yet. The student preparing calmly over six months is explicitly out of scope – designing for both is a reliable way to design for neither.

No users. So the limits are part of the deliverable.

Every number carries its source. Where there is no link, the statement is labelled as an estimate or an assumption.

A growing market with a hole in the middle.

Between free tools and the cheapest paid course there is nothing – and that gap is exactly where someone with three weeks left is standing.

For a product priced at $10–40 the middle of this ladder is empty space. The person with nineteen days left is not going to buy a three-month course – and nothing else is on the shelf.

Six products, one question each.

Not a feature comparison. For every product: what behaviour is it designed for, and what breaks when the user has three weeks?

Product	Price & horizon	What breaks in a short window
Khan Academy	Free · 8–12 weeks	The plan is skill-shaped, not time-shaped. The volume overwhelms without outside guidance.
Bluebook	Free · 2h14m tests	The smallest unit of work is a two-hour test. Half of test takers complete none.
UWorld	$249–300 · 30–360 days	The shortest tier is still 30 days. You buy access time, not a result by a date.
Princeton Review	$299 · 365 days	Nothing tells you which of the twelve hours of video matters when you have twenty.
Kaplan	$249–799 · 6 months	The pace belongs to a class schedule, not to your exam date.
AI tutors	$49–70 / month	Content-first and priced per month. A three-week user pays for a full one and still decides what to study.

The pricing proves it: tiers of 30, 90, 180 and 365 days. And every one assumes the hard part is knowing what to study. For this segment the hard part is opening the app.

Nine threads, read the way transcripts are read.

Threads were found using the segment's own words – “two weeks left”, “haven't started”, “is it too late”, “I can't start” – then read for what people say before anyone asks them a question.

Pattern 01 – they open with a verdict on themselves

Steel's two strongest predictors of delay – task aversiveness and low self-efficacy – show up in the opening line. A product that greets this person with a deficit (“you are 240 points short”) feeds the exact mechanism keeping them out of the app.

Pattern 02 – the most repeated question is a forecast request

People ask the internet for a forecast because no product gives them one. A paid tutor answers with three inputs – target, baseline, days remaining – in the first two minutes of a session that costs $45–100 an hour. That is the screen, requested unprompted.

Pattern 03 – they think in days left, never in curriculum

Every post is titled with a time budget: “I have the SAT in two weeks and I haven't started anything.” Not one asks which topics the SAT covers. The unit people organise around is the one no product is built around.

Pattern 04 – almost nobody in this corpus chose to be here

School Day puts the date on the calendar, a reinstated requirement puts it back, a deadline abroad forces it, parents insist. In every version the decision belonged to someone else – which is why a product designed for someone who decided to prepare misses them entirely. It also complicates payment: the person who feels the urgency and the person holding the card are often not the same.

Mechanisms, each translated into a decision.

With no access to users, behaviour is explained from research – and every mechanism is turned into an interface decision, or it is useless.

Segments by behaviour, not demographics.

The job, stated

Note what is absent from that job: no wish to master the material, no interest in a curriculum, no goal of becoming a better student. The job is to convert whatever time is left into the best available outcome.

Three insights

Seven beliefs, written down before the research started.

Scored by importance and uncertainty, with only the top group worked on. Two of them moved the wrong way – and stayed in the report.

Keeping A4 and A5 in the deck is the point. A research document that only confirms its author is a pitch, not research.

Six rules – each one stating what it rejects.

A principle that rejects nothing is decoration, not a tool.

What we believe → what we build → what we expect.

Today the entry ritual is a 2h14m practice test, then a verdict, then a paywall. For someone with twenty usable hours that is a tenth of everything, spent on measurement before a single point is earned.

Nothing here is assembled from taste. The cost of the old entry comes from College Board, where the gain lives comes from College Board, what the user asks for comes from the threads, the formula for answering it comes from a tutor in those threads, what nobody on the market does comes from six teardowns, and what a minute of practice is worth comes from retrieval research. The bet is only in putting them together.

North star: points gained per hour spent

The only metric that survives this segment's constraint. Someone with twenty hours does not care about engagement, mastery or habit – they care what those twenty hours convert into. Deliberately not counted as success: daily active users, streak length, minutes in app. Those measure the product this segment already abandoned.

What this research did not settle.

The product never launched. There are no conversions and no A/B tests, and there cannot be. What can be shown is principles, a concept, and decision rules written in advance.

Decision rules, written before the data

1. If the reachable-range framing does not beat deficit framing on continuation, Principle 04 is dropped rather than defended.
2. If deadline pricing converts no better than monthly, the pricing insight was a designer's idea, not a user need.
3. If predicted ranges miss real scores badly, the forecast is removed entirely – a wrong number is worse than no number.

Every number in this case has a link.

Where there is no link below, the statement in the case is an estimate or an assumption – and it is labelled as one.