The Derelict Starship Problem: Why Good Sci-Fi Games Need Real Stakes

There's a moment most AI sci-fi games die, and it happens around turn four.

You've prompted the model. It has dutifully described the air — "stale, metallic, faintly burned" — and the hum of a reactor that shouldn't still be running. It has told you there's a corpse in the corridor and a console blinking red. Then it asks, in that infinitely patient way models do: What do you want to do?

And you sit there, cursor blinking back at the blinking console, and realize you don't care.

That's the derelict starship problem. The setting is vivid. The prose is competent. The model knows what a sci-fi game sounds like because it has read ten thousand of them. But nothing is actually at stake. You could type "I examine the corpse" or "I leave the ship" or "I start singing the anthem of Proxima Centauri" and the story would adjust with equal grace. The model is a mirror; you are standing in front of it asking it to be a cliff.

Games need cliffs. And the reason so many AI-generated sci-fi games feel flat isn't that language models can't describe things. It's that the games don't give you anything to lose.

What "stakes" actually means in an AI game

Stakes are not "the stakes are high." Stakes are a specific mechanical thing: a value the player cares about that can move up or down based on their choices, and whose final state matters at the end.

A chess game has stakes because pieces leave the board and don't come back. A D&D campaign has stakes because your wizard took six months to level and dying means making a new sheet. A good novel has stakes because you've spent two hundred pages learning to like somebody who might die on page three-oh-four.

AI text games, by default, have none of this. Everything the player "has" exists only in the model's context window. The model can resurrect the corpse, unburn the reactor, and uncommit the murder in a single hand-wave if you ask nicely. The medium itself fights against permanence. Which means the game has to smuggle stakes in deliberately, or the whole thing dissolves into collaborative flavor text.

Three a-gnt games do this well, in three very different ways. Looking at what each does — and what a flat game leaves out — is the fastest way to understand the problem.

🚀Into the Derelict: stakes through resource math

🚀Into the Derelict is the one that taught me the derelict starship problem has a solution.

It opens the same way most of its cousins do: you've boarded an abandoned ship, something is wrong, the lights are flickering. But it does one specific thing different in the setup. It gives you a resource counter. Oxygen: 94 minutes. Battery: 68%. Radiation dose: accumulating. These are not decorative. The prompt explicitly tells the model to decrement them every turn, and to kill the player when they hit zero.

That sounds crude. It isn't. It's the entire difference between a mood piece and a game.

Because now when the model describes a long corridor with a sealed door at the end, you don't just drift down it. You count minutes. You think: do I spend six minutes forcing this door or go around? Around is eleven minutes. Around costs more oxygen but avoids the radiation pocket near the reactor. Which one? You have to actually choose. The model is no longer a mirror — it's a clock, and the clock is running.

The craft observation: stakes don't have to be dramatic. They have to be countable. A number the player can watch change is worth more than a thousand words of atmosphere. Atmosphere enriches stakes; it cannot substitute for them.

🔍Quantum Detective: stakes through unreliable knowledge

🔍Quantum Detective solves the problem from the opposite direction. It has no hit points, no oxygen, no death state at all. It's a mystery. You're investigating an impossible murder on a research station that's running quantum experiments, and the only thing you can lose is being right.

What makes it work is that the game tracks a hidden truth. At the start, the model privately decides which of several possible solutions is correct — a specific perpetrator, a specific method, a specific motive — and then refuses to change its mind. You can interview witnesses, inspect the scene, get distracted by red herrings, pursue theories. Every piece of evidence the model reveals is either genuinely consistent with the answer it committed to at turn zero, or a deliberate, tracked misdirection.

The stakes are epistemic. You either figure out what happened or you don't. The game has an ending condition (you accuse someone), a win state (you're right), and a loss state (you're wrong and the real killer gets away in the prompt's epilogue). The number you're tracking isn't a counter — it's your confidence, and it's in your own head.

This is a trick that a lot of AI mystery prompts miss. They let the model improvise the solution as it goes, which means there's never a moment of genuine "aha" — the truth is whatever makes the player's last guess feel dramatic. That's fine as theater. It isn't a game. A game requires the answer to exist before you start looking for it.

🌀Escape the Event Horizon: stakes through time pressure you can feel

The third example, 🌀Escape the Event Horizon, uses a device that's almost embarrassingly simple: the game is structured in rounds with a declining turn budget. You have twelve turns to get off the ship before it crosses the event horizon. Eleven. Ten. Every action you take — every examined panel, every conversation with the injured engineer, every argument with the AI — burns a turn, and the model is instructed to refuse any "free" action.

That constraint forces a conversation most AI games never have. You cannot explore everything. You have to decide what to ignore. The engineer is dying and might have the override code, but the comm array might still reach a rescue ship, and the escape pod might or might not have fuel, and you have nine turns. Pick two.

What stakes do here is produce regret. When you finish an 🌀Escape the Event Horizon run, you remember the specific thing you chose not to do. You remember leaving the engineer behind. You remember not checking the pod's fuel. The game is memorable because the game took something from you — not hit points, but attention. A finite amount of attention is an extraordinarily cheap way to create meaning.

Compare that to a game where you can try every door, read every log, ask every question, and eventually arrive at the same ending whichever route you took. That isn't a game. That's an audiobook with a choose-your-own-adventure veneer.

The failure mode: the infinite-sandbox problem

If you want to see what "no stakes" looks like, spin up a generic "you are a starship captain, what do you do?" prompt and play for ten minutes. The model will cheerfully generate whatever you ask for. Encounter a pirate? You win. Or you lose and somehow survive. Or you negotiate and get a valuable ally. Everything is possible because nothing is refused. The sandbox is infinite and therefore weightless.

This happens even with genuinely clever openings. 🎮Asteroid Field Pilot is interesting here because earlier versions of it had exactly this problem — a beautiful setup, a great voice, and an impossible-to-lose core loop. The current version solves it with what I'd call narrative hull integrity: a hidden damage counter that persists across turns, and a rule that certain maneuvers automatically damage the ship. You can be reckless. Recklessness has a cost. The cost is tracked. When the ship explodes, it explodes because of choices you actually made, not because the model felt the scene needed tension.

Good sci-fi games aren't lenient. They're consistent. Consistency is the stake, because it means every choice you made up to this point was real.

Game masters versus puzzle boxes

There's a second kind of AI sci-fi game that works differently: the persistent game master. The standout in the catalog is 🎲Game Master Galactica, which isn't a single game prompt but a long-running role. You tell it what kind of campaign you want; it runs it for you across sessions; it remembers what happened two weeks ago.

The interesting thing about 🎲Game Master Galactica from a stakes perspective is that it externalizes memory. Instead of relying on the model's context window (which forgets) it instructs the model to keep a structured session log — characters, locations, unresolved threads, debts owed, promises made. That log is the stakes. When your engineer owes a debt to a Hutt-equivalent in session three, that debt is still there in session seven, and the game master will bring it back without you asking.

This is the hardest kind of stake to engineer in an AI game because it requires the tool to work against the model's natural tendency to forget. But it's also the kind that produces the deepest play, because it lets a player form an attachment to a character across real time. You can't love a starship captain you met four turns ago. You can love one you've been playing for six weeks.

Three stakes-design principles, stolen from tabletop

If I had to distill what the working a-gnt sci-fi games have in common — what separates them from the flat ones — it's three principles that tabletop designers have known forever:

First, something must be countable. A number, a clock, a list of names. Something the player can see tick down. Atmosphere is seasoning; the counter is the meal.

Second, something must be refused. The model must be instructed to say no. No, you can't talk your way past the airlock without the code. No, the engineer is dead and stays dead. No, you used your last medkit in turn four. A game where nothing is ever refused is a game where nothing is ever earned.

Third, something must persist. Whether it's a resource, a piece of knowledge, a relationship, or a hidden truth the model committed to at turn zero, there must be a fact about the game world that was true before the player's current message and will still be true after. Otherwise every turn is the first turn and nothing adds up to a story.

All three of the games I named above do all three things. The flat ones — the drifty mood-piece ones you forget before the tab closes — usually do zero.

What a better sci-fi game looks like from the inside

Close your eyes and picture the moment a good one lands. You're playing 🚀Into the Derelict. You have fourteen minutes of oxygen. There are two doors. Behind door A is probably the reactor control you need. Behind door B, a voice you haven't identified is calling for help in a language the translator is only half-rendering. Door A saves you. Door B saves someone else, maybe, if they're real, and costs you four minutes you probably can't spare.

You sit with the choice for a second longer than you'd sit with any decision in a game without stakes. That second is the whole thing. That second is what every good sci-fi game is trying to produce. Everything else — the prose, the atmosphere, the ozone and the flickering lights — exists to make that second feel heavy.

A flat game never produces that second. You read the prompt, you shrug, you type the first thing that comes to mind, and you close the tab. A game with real stakes produces it once a turn.

The craft handoff

If you're building a sci-fi game on a-gnt — or anywhere else — here's the specific test to run before you ship it. Play it for ten turns. Then ask: could I have typed almost anything and gotten an equally valid result? If the answer is yes, the game doesn't have stakes yet. Go back and add a counter, a refusal, or a fact that persists. Any of the three will do. All three is better.

The derelict starship problem isn't that AI can't describe a derelict starship. It's that a vividly described empty room is still an empty room. Stakes are what turn the room into a choice. Choices are what turn a choice into a memory. And memory is what turns a game into something someone actually wants to play twice.

Fourteen minutes of oxygen. Two doors. That's the whole craft, right there.