22 two Case Study

22 two is an AI interview coach. It runs a mock interview for a role, scores the answers, and compares your CV against the job description you are actually applying to. Free covers the interview side. Pro covers the CV side.

I designed the feedback system, the CV-to-JD analysis behind the paywall, the upgrade moment between them, and the billing screen that makes that moment defensible.

The product could already run an interview and grade it. What it could not do was tell you anything. Score, transcript and observations landed on one page at the same weight, so people read the number, felt something, and closed the tab.

A grade is not coaching. The expensive work was already done and then handed back as raw material for the user to interpret themselves.

I looked at what the category does with a paywall. There are three moves: meter every feature, blur the paid screen, or show a number and charge for the reason. All three treat the free tier as a trailer.

The pattern underneath is that these products sell practice, and practice is a habit. Habits do not survive being rationed. That reframed the commercial question from how much to hold back to which whole thing to give away.

Sessions with the founders to settle one thing before anything was drawn: what free is allowed to be. Growth wanted the answer-by-answer breakdown capped at three questions, on the reasonable grounds that it is the most expensive screen we give away.

The decision that set the tone: free gets the whole interview feedback system, uncapped, and the paywall moves to a different job rather than sitting partway through the same one. Practice is how you improve over months. A CV gap analysis is what you need the night before applications close. Deadlines are what people pay for.

I watched people read their own feedback. The tell was where they stopped. Everyone read the top score, most read the skill breakdown, and almost nobody reached the transcript, because a transcript asks you to do the analysis yourself.

Two things came out of it. People want the reason before the remedy: hearing what a strong answer sounds like is useless until you know why yours was not one. And nobody reads six questions carefully, they scan for the two that went badly. That made scannability a requirement of the depth layer, not a finish on it.

The spine is a session. Run the interview, get a score you can act on in thirty seconds, go deeper on the two answers that cost you marks, then find out whether your CV would get you into the room at all.

The overview and the breakdown are two tabs of one free screen, not two products. The overview answers how did I do. The breakdown answers what do I say instead. Splitting them is what lets six sub-scores exist without the page turning into a dashboard nobody opens.

Every answer card runs the same three moves in the same order: what you said, what it demonstrated, a stronger version. The repetition is the feature. Once you have read one card you can skim five.

Two label sets carry state. Clearly demonstrated against Expected but missing on answers; Strong, Needs work, Weak and Complete on CV sections. One decision was made against instinct: What worked is a full section, not a courtesy line. A scoring product that shows only deficits teaches people to dread opening it, and the people using this one are nervous before they arrive.

The paywall took the most rounds. I drew three: the whole analysis softly blurred, the top-line numbers exposed with every explanation locked, and one real gap from the user's own CV solved all the way through.

The middle one is what a growth team picks, and it is the one that makes people resent you. Telling someone they are a 63% match and charging them to find out why is a hostage negotiation. I recommended the third. It costs the most to give away and it is the only one that proves the product works rather than proving it exists, someone who has read one finished gap knows what the other five are worth.

Interview feedback opens on one score, four delivery measurements (speaking pace, filler words, clarity, talk time) and six skills scored separately, because you scored 76 is a verdict and your STAR usage is 62 is an instruction. The second tab carries the per-question cards.

The paid CV screen is specific about one CV and one job description: match score, keyword coverage, ATS compatibility, six named gaps ranked by how much closing them moves the score, and before-and-after rewrites of the user's own lines. Every one is a suggested edit. Applying opens a diff, never a file that changed underneath them.

Plan and billing is the quiet half. Cancel plan sits as a visible link at the top, not buried under Account and not three confirmations deep. Trust is not a property of the screen where you ask for the card. It is a property of the screen where people find out how hard leaving is.

Four core screens: interview feedback with its overview and answer-by-answer tabs, the paid CV-to-job-description analysis, and plan and billing. Three paywall concepts drawn and argued, and one recommended, the one that solves a real gap from the user's own CV all the way through instead of blurring the answer.

The recommendation stood: an uncapped free tier, with the paywall moved onto a different job rather than sitting partway through the same one. I have no conversion data to show for it. The comparison that would settle the argument is free-to-paid against the blurred version, cohorted either side of the switch, and it is the number I would have asked for first.

Recommending the third paywall was the highest-leverage decision in the project. It moved the ask from withholding evidence to supplying it, and it turned the quality of the free tier into a commercial asset instead of a cost to be managed.

What I would do differently: the mobile layout for the CV screen. Three dense tables and a before-and-after column set are a desktop design I compressed rather than rethought, and compressing is not designing. The narrow case is the harder constraint and it should have led.