ten prototypes · four synthetic testers · Aug 21 2026
Read this first. The four testers are AI-simulated teenagers, not real students. They looked at real screenshots of the real prototypes and reasoned about them. That can tell you whether something is legible, whether a reward moment is visible, and how the ten compare to each other. It cannot tell you whether anyone comes back tomorrow — and a 90-second test structurally favours the arcade-y ideas over the come-back-tomorrow ones. Treat this as a filter for which directions are alive, not as a verdict on which to build.
| # | Game | Pts | Places | What it was testing |
|---|---|---|---|---|
| 1 | BAIT | 32 | 1,2,3,6 | an inverted relationship to the material |
| 1 | LOADOUT | 32 | 1,1,4,6 | does a meta-layer lift the same drill |
| 3 | PR | 30 | 1,3,4,6 | competition with no opponent |
| 4 | BANK IT | 29 | 2,4,4,5 | greed — the only route with no clock |
| 5 | TODAY'S TEN | 23 | 2,3,7,9 | closure |
| 6 | BRACKET | 20 | 2,6,8,8 | a rank moving under you |
| 7 | NUMBER | 16 | 5,5,8,10 | does a falling number feel like stakes |
| 8 | SURE? | 14 | 3,8,9,10 | calibration as a ranked axis |
| 9 | TEST DAY | 12 | 7,7,9,9 | section-cutoff pacing |
| 9 | HOLD | 12 | 5,7,10,10 | math as a weapon (PvE co-op) |
10 points for a first place, 1 for last, summed across four testers. The spread matters more than the total — LOADOUT took two firsts and a sixth. Nothing here was universally liked.
Every tester was asked separately which was most fun now and which they'd still open in two weeks. Three of four gave different answers, unprompted.
"Most fun right now is #9 — the only one I wanted to immediately replay. But LOADOUT is a run-based game with eight perks and no progression between runs, so in two weeks I'll have found the good build and there'll be no reason to open it. #2 is the one that survives because it doesn't depend on being fun. It depends on a number I don't want to reset. That's a different mechanism and it's a stronger one."
— Cole, 16, completionist
TODAY'S TEN ranked 5th for fun and won the two-week question. That is the whole lesson of this exercise in one line, and it is exactly what the format was predicted to get wrong.
Two testers independently named the identical moment, unprompted — and neither ranked that game first:
"In #5 PR, the track flipping red → green the second I pull ahead of the ghost. Not the end screen, not the PR banner. The flip, mid-run, while I'm still going. That's the only second in all ten where my body did anything."
— DeShawn, 17
"It told me I was winning before the run was over. Every other game in this batch makes you wait until the end to find out how you did."
— Cole, 16
Take the mid-run feedback flip and put it inside whatever we build. It is not specific to PR — it's a property (tell the player they're winning during, not after) that most of the other nine lack.
HOLD is your idea, built the way it survives the objections: no human opponent to humiliate anyone, your math is the weapon, an ally instead of an enemy. It finished last-equal. Two of four named it the one they'd be embarrassed to be seen playing.
The honest part: I expected it might lose because it has no artist, which would have been a false negative. That is not what happened. All four rejected it for the same structural reason, and they found it independently:
"It's a tower defense where the tower is math, which means the game part and the learning part have nothing to do with each other — the orange squares are just a progress bar wearing a costume."
— Cole, 16
"A fake friend generating fake need so I feel important for solving 2x − 5 = 11."
— Maya, 16
The distinction is worth holding onto: the four games that won all have the math inside the mechanic — in BAIT the math is the puzzle, in BANK IT the math is the risk. In HOLD the math is a button you press to make an unrelated animation happen. That's a fixable property, not a dead idea — but combat framing alone doesn't fix it.
One thing DeShawn said that's worth keeping from the wreckage:
"I'd rather have a fake teammate than a real opponent, because a fake one can't see me miss. That part's actually right."
— DeShawn, 17, projected ACT 19
Two testers independently attacked the questions, not the games:
"The math is fifth grade. 3² + 4². −5 + 12. I'm projected 30, I don't get those wrong when I'm half asleep. Every one of these mechanics is a wrapper around a question. A ghost race where I can't lose isn't a ghost race."
— Maya, 16
That's a fair hit on the shared item pool, which skews easy. It means the difficulty ceiling is too low to generate real tension for a strong student — so every ranking here is measured against questions that were too easy, and the timed routes are penalised most.