The game runs 21 turns. At the end, it produces a document.
The document is a description — your dominant trait, a structural blind spot, a prediction about where you're going. Players report it's more accurate than what they'd have written on their own. NPS 9 from people who don't finish. Emails back saying it gave them better perspective than months of ChatGPT and journaling combined.
The first instinct is to call this AI therapy — Betterhelp territory, Talkspace territory. That instinct points at something real: the conversations go deep, the content is personal, players share things they haven't said out loud before. The instinct is correct that this is a mental infrastructure product.
The analogy that actually fits is Replit.
Replit didn't set out to make coding easier. It made execution infrastructure. The work happened inside Replit, so the corpus of what people built there became the primary asset — not the IDE itself. What Replit owns is the record of how developers actually think, iterate, and solve. That corpus compounds. Each project adds signal about what real development looks like under real conditions.
The session is the door. The corpus is the house. The individual session has value. What accumulates is what matters.
The data problem in AI isn't quantity. It's depth.
Most behavioral data comes from zero-stakes conditions — the least interesting version of a person. A ChatGPT conversation costs nothing to abandon. A journaling app doesn't push back. The version of yourself that shows up when nothing is on the line is a lower-resolution version.
Someone who paid to be there, inside a conversation that's actively calibrating to them, under conditions they chose — that person generates a different kind of signal. The self they're describing is the one they actually live with.
The intersection of willing, filtered, high-stakes, and longitudinal is essentially empty right now. Therapy apps have enrollment but the conversations are episodic and unstructured, optimized for symptom relief rather than self-model precision. Social platforms have scale but not depth — what you post is performed, not disclosed. AI companies have massive usage data but it's passive, opt-out-by-default, zero-stake behavioral trace. None of these produce what a person writes when they've committed, know they're being read carefully, and are answering questions they didn't expect to face.
The corpus compounds because each new session enters a record that's already seen every prior session. Accuracy grows through comparison — the system gets better at recognizing what a specific kind of person looks like at specific inflection points in their development. That calibration can't be built retrospectively. It has to be earned in real time, from real people, with real stakes.
The window is short — not because the technology is early (it isn't), but because the people worth collecting this from are increasingly selective. High-agency operators already have a dozen AI tools competing for their attention. The ones still willing to go deep with something new, to pay, to answer uncomfortable questions about themselves — that population shrinks as the space gets noisier. The filter that makes this data valuable is also the thing that gets harder to run as time passes.
Being first here isn't a virtue claim. It's a description of where we are in the clock.
The reader who recognized themselves somewhere in this post is the person the game was built for. Not as a demographic description. As a design constraint — the product was designed to find you, and finding it means you were findable.
The filter has been running this whole post.
Go see what it found.
P.S. If you've played: your session is already in the corpus. If you haven't: it starts at the beginning.