The one-shot game that worked, and the one that didn't

I was messing around with Sora, trying to find prompts that produced something interesting, and I asked it for a skateboarding wizard in Venice Beach casting fireballs at homeless people.

It did a great job. And the image stuck with me for weeks — it just looked fun. So eventually I built the game.

You can play it here. It’s about fifteen minutes. Keyboard and mouse, sorry.

What it is

You’re Kai, a skateboarding wizard with a 2 AM open-mic slot and no poem written. You have until then to skate Venice and Santa Monica, collect twelve pieces of inspiration, and survive the boardwalk.

Grind on anything with an edge — curbs, cars, benches, string lights strung between buildings, the VENICE sign. Fireball the fiends. Refill your mana at food shops, because that’s the only place it refills. Buy drip with the cash the fiends drop, and then every piece of clothing earns you exactly one unique compliment from a stranger on the street, which raises your EGO, which maxes your stats and makes your fireballs bigger.

Kai hates this. He says so.

That last mechanic is lifted, with affection, from Bum Bag Bangin’ by SeaDads — a Ludum Dare 40 game I’ve loved for years, where you’re a guy walking to work in a genuinely excellent bum bag and the compliments are the hazard. Too many and your head swells. I inverted it: here the compliments buff you, and the joke is that your wizard resents being helped by them.

Everything you see is procedural. No downloaded models, no textures, no asset packs — every building, sign, and facade is painted to a canvas at runtime. All 76 audio assets (51 voice lines, 22 sound effects, 3 music tracks) came out of the ElevenLabs API.

The prompt was the whole thing

I’ve seen people describe one-shot game generation as a kind of magic trick. It isn’t. The result was good because the prompt was long and specific, and for no other reason.

I didn’t ask for “a skateboarding game.” I named every mechanic. Skating, grinding on rails and sidewalks and cars, wall-jumping off buildings. I specified the restaurants you could go into — and I specified the joke that happens at each one. I specified the clothing shops, and the compliment loop, and that the character should comment on his own ego inflating.

Then Fable 5 wrote the actual lines, and this is the part I didn’t expect: it is genuinely funny. The voice lines land. The shop scenes land. I gave it the shape of every joke and it filled them in better than I would have.

I don’t think that contradicts what I found building BLACKWATER — that the model resists creative writing. It’s the same finding with the constraint changed. Here every piece of writing was small, bounded, and pinned to a mechanic I’d already specified. That’s not a long-horizon creative task. That’s filling in a form, and it’s excellent at filling in forms.

Then I got greedy

I was so happy with how this came out that I ran it again, bigger.

Same model, same settings — Fable 5, ultracode, maximum reasoning. Same prompt, too, with one change: I added more detail. The small game only needed downtown and the beach. The big one asked for six districts, including a suburb, the city drainage system, and a park, and I described what I wanted in each of them at much greater length.

I got a world about three times the size.

The map had holes — buildings that should have been there and weren’t. Clothing lines strung between nothing, floating in mid-air. The park was giant and empty: no trees, no benches, no paths.

That much I could explain to myself. More ground to cover, less care per square foot. Fine.

Here’s what I didn’t expect. The parts I hadn’t changed got worse too.

Kai was less likeable. The store owners had no good jokes. The cutscene camera work was far worse. A pile of small touches that made the first one feel finished were simply missing, and I couldn’t point at the line in the prompt that had asked for any of it to change, because there wasn’t one.

The clearest one is the friend scene. In the small game, when you stop to talk to your friend, the game pauses so you can watch it. I never asked for that. It just did it. In the big game there’s no pause — but you also can’t steer during the conversation, so Kai keeps his momentum, rolls off the edge of the screen, and the whole thing plays out with him somewhere off camera.

Nothing in the prompt about that scene had changed. It got worse anyway, while the model was busy somewhere else in the city.

The obvious lesson from a bad generation is that your prompt wasn’t specific enough. I believed that. This run disproves it: I wrote a more specific prompt and got a worse game, including in the parts both prompts described the same way.

So what a one-shot build spends isn’t specificity. It’s care, and the care is global. Adding a district doesn’t cost you a district; it costs you everywhere. I didn’t buy a bad park and keep everything else intact. I paid for that park out of Kai’s dialogue, the shop scenes, and the camera.

It isn’t a ceiling on what the model can do, either. The small run wrote dialogue funnier than what I’d have written and I hadn’t specified a line of it. There was room to spare. Nothing about the big run says it couldn’t have done that again. It says it was busy.

That failure is what sent me down the road that became BLACKWATER — built iteratively, with developer tools and tests requested first, so I could see and fix what the model couldn’t. Those are the same project in a sense. This one is the control group.

The size rule

So here’s the thing I’d tell anyone trying this.

One-shot generation works, and the constraint is scope, not ambition. I still think that’s right. I just had the mechanism backwards. Scope isn’t a cap on how much you get out. It’s a setting on how much care every part of the build gets, and you change it for the whole build at once.

If you want a genuinely good game out of a single prompt, it has to be about this big — roughly a fifteen-minute experience, two districts, a handful of interlocking systems, one joke per location. At that size you can be specific about everything, and specificity is the whole input. It’s the whole input only because at that size it’s still affordable.

The first things to go are the ones you never asked for, and in the small game those are a lot of what makes it good. I couldn’t have listed them in advance either. I didn’t know they were there until they were missing.

Go bigger and you don’t get a bigger version of this. You get slop, and not only in the parts you added. So the way to check whether you’ve overshot isn’t to inspect the new stuff. It’s to go look at something you didn’t touch and see whether it still works.

Also: give it the ElevenLabs API. That’s not scope-dependent. That’s just always right.

Why I keep doing this

Honestly, mostly because it’s fun, and because there’s something specific and good about watching an idea you’ve been carrying around turn into a thing you can actually load and play.

This one took an evening. About four and a half hours from the first file to the last, which is longer than the hour I remembered and still absurdly short for what came out.

I’d like to do more at this size.