选择你的灭亡

I pick up a butter knife from our picnic and stab the dragon in the eye.

Your AI storyteller is a coward

Stories and games are boring without stakes. Death doesn't need to stalk every page, but in a combat-heavy or survival-focused story, fatal situations will come up sometimes, and the story needs to handle them consistently! Our story harness allows the player's choices to meaningfully shape the story, which means those player choices have to have consequences — even the "bad" ones.

Let’s consider a simple AI DM, such as you might experience in D&D-style tools like SillyTavern or AI Realm, or even setting up your own custom prompt in LLM chat tools like Claude, ChatGPT, or Gemini. If you tell your AI DM that you attack the dragon with a butter knife, you will almost inevitably emerge the unlikely victor. Probably without even a meaningful sacrifice, which would be one way of preventing a cheap underdog victory (see, for example, Robin McKinley's The Hero and the Crown).

A good human DM might laugh and say "You can certainly try," then have your character swatted aside with some bruised ribs and a chance to think of a better plan. A character who insists on this course of action again and again will probably end up as crispy dragon dinner.

The AI narrator, on the other hand, is going to cheerfully sand the edges off of everything unless you specifically tell it to give you a consequence, and policing yourself isn't nearly as fun as being surprised. And gameplay without boundaries, where the story says yes to everything, gets dissatisfying after a few minutes. You zoom straight from fighting the village rats to the classic defeat-the-gods-and-become-one plot, and now you're bored. Things keep happening, one after the other, but it feels infinitely mid.

Handling a moment like the butter knife dragon attack requires good judgement, the ability to balance player agency with good narrative and earned moments, and the possibility for things to go properly, catastrophically wrong. That doesn't mean you can't have your moment of utensil-derived glory! It's fine if a scrappy butter-knife victory is appropriate to the arc and the tone of the setting, with or without permanent damage to your character. But that only works if this kind of gambit doesn't work every single time. The game has to be able to tell you "no" sometimes.

Death is the ultimate "no", way out at the far end of failures.

Modern model training sets models up for this kind of cowardice. Because of the chatbot explosion, post-training rewards cooperative, helpful behavior, and the far end of that is sycophancy. Yes, you're absolutely right! Meting out real consequences, even in a story, cuts against that training. Without the player constantly steering the model back into character, it will back out of even the mildest conflict or cost. That's annoying behavior in a coding assistant, let alone a storytelling DM. The flip side is that you can't tune a model to one perfect level of punishment, either. The right amount of "no" depends on the story, the moment, and the player.

The pitfalls of peril and death go beyond plot armor, though. There's the usual AI storytelling challenge of amnesia, like a serious injury that gets dropped between chapters, or the enemy who beat you and never comes up again, or seems to have forgotten that they kicked your ass last time. The narration does whatever is locally plausible next, and the most recent turn tends to outweigh everything that came before it.

Five bad habits of AI telling a story

We'll refer back to these throughout this series. Each one is tricky because, annoying as they are, you’ll get problems in the opposite direction if you push naively against the tendency.

Bad Habit	What Happens
Sycophancy	It reads your action as a request, and requests get granted (But go too far the other way, and you’ll end up with an inflexible narrator that spoils the fun and ruins the "collaborative" part of collaborative storytelling much more quickly than a too-permissive one).
Local plausibility	It writes the most plausible next sentence or action given what's in front of it, even when that doesn’t make sense in the context of the story and characterization so far.
Recency	The last turn of events has disproportionate influence on what comes next; this makes structurally pleasing callbacks difficult, and can also lead to a local hyperfocus on an otherwise throwaway moment. A classic example is a character who is annoyed at you escalating into outright hostility over the long term for no particular reason, just that their most recent interaction with you was a conflict (And yet the most recent part of the story is generally the most relevant, on average, so recency should win sometimes).
Regression to the genre mean	In the absence of specific choices at any level, it tends to play toward the most common trope in the genre (But familiar patterns are also what make tropes pay off and satisfy in a big way, and variation over time is what keeps stories interesting).
Eagerness to resolve tension	An LLM loves wrapping things up with a bow, a concluding sentence, and probably a moral. It feels trite and boring, and also, bad things need to happen for stories to progress (Yet in many conflicts between characters, and especially if you push the story toward tension, it will tend toward infinite sabre-rattling…because, again, AI DMs are cowards).

It helps to think about what death actually does to the player's choices. In most games, dying resets you to a place with no choices at all (game over) or with almost the full set of choices restored to you (reload the save, or start over from the beginning). In a written story, death is often at least a plausible risk, but sometimes it’s also a transition that changes which choices you have. The protagonist of The Crow doesn’t end their story by death, but who they become and what choices they have to make are fundamentally changed by their post-death return.

A lot of narrative games are sold on the promise of infinite choices, but as many character chats and procedural generation explorations have proven, infinite choice isn’t actually all that fun. The fun comes from constraint, from making a choice that has risk and costs. But it’s hard to lay the constraints out yourself and still play with them, and that’s what a good DM should manage on your behalf. This is where that "good judgment" comes back into play in our systems. Where can a player demonstrate agency and build towards earned, rewarding narrative moments? What are the (limited!) number of ways that things might go properly, catastrophically wrong?

This is where our narrative world model really shines: in its world context and its history, and in its playbooks. So what comes next after death? With our agentic story harness, you get the only correct answer: *it depends. *

Failing forward, even in death

If death is the extreme end of failure, then how do you fail forward from it? And what does "it depends" actually depend upon?

Tabletop players know the phrase fail forward: a failure should move the story somewhere rather than stop it; it should always be interesting, and in the right circumstances that can even motivate the player to make “bad” choices for the sake of an interesting story. We hold death to the same standard. For that choice and failure to matter, the consequences should move the story somewhere new, and it should be folded into the premise of whatever comes next at the appropriate scale, whether that's the next scene, the next chapter, or the whole storyline. At Hidden Door, we have a set of playbooks devoted to exactly this pattern.

Stories and games both run on rules about what death and defeat mean, and the best ones play with those rules. Every story has defaults. Death ends a character's story, and being knocked out means you failed. Stories get a lot of their power from bending those defaults on purpose. In The Crow, death is where the protagonist's story begins. In Hades, death is mostly a different realm, though you can die within that too (you can imagine that the structured backend of death becomes complicated quickly, when accounting for all these kinds of death!). In The Wizard of Oz, getting knocked out is the thing that transitions you out of ordinary life and into the fantastical realm. You have to adhere to the rules well in order to break them effectively.

So choosing how a story handles death is a storytelling decision before it's a technical one. It sets the rules, and it decides which rules a world is allowed to bend.

How we do it

Broadly, we have a few different settings for death. Each one is right in a different context, and together they let creators and players set up the social contract before anyone rolls dice.

No death. Suitable for a cozy mystery, a children's adventure, a comedy. In mechanical terms, falling to zero energy makes the character unconscious, and conditions that would normally be fatal, like losing a duel, render you unconscious instead. The scene ends, and your next scene opens with some kind of recovery or convalescence. If you get nailed by lasers during the heist, you might wake up in jail or in the billionaire's dungeon, but you aren't dead. We use the superpowers of LLMs to shape that into a coherent narrative. It's plot armor, but on purpose!

Normal death. The default, and suitable for most genres. Zero energy works the same as in a world without death, but characters can also acquire fatal conditions. Whatever goal you were pursuing, the scene ends in failure. If the protagonist is fatally wounded, the player gets to pick a more significant recovery scene: you live, but the storyline you were on is steered in a different direction for good. In worlds where the laws of physics and magic permit it, you might genuinely die and be brought back. The player could even switch characters, with a playbook for reviving your only-mostly-dead friend.

Permadeath. For players and worlds that want real stakes. Death gets a real ending, written in the world's voice, and the character's story is over. The larger story can go on with someone else, which unlocks some really neat playbooks: a vengeance quest led by the people you left behind, or picking up the dynastic torch. Sometimes the fun is playing the story like a director, rather than maintaining one single character’s viewpoint.

What building it taught us

The world's rules decide what death means. The narration writes it, and never gets to decide whether you live. Left to their own devices, most models will yield to sycophancy, eagerness to resolve tension, and regression to the genre mean. Models are good at generating words. They are much less good at storytelling, which is the part where you decide what happens. So we split the job. Plain mechanics decide whether the character is down and what this world allows next. A small set of hand-written recovery playbooks decides the shape of what comes next: carried to shelter by whoever was there, waking up in whatever passes for a healer's tent in this world, held by the one who beat you, or the others regrouping without you. The model's job is to pick the one that fits what just happened and fill it in: who was there, how much time passed, why the danger is gone for now.

You still have to tune the mechanics that enforce the rules. For a while, characters in non-permadeath worlds recovered to 1 HP at the start of the next scene. That was a bit too easy to lose, and the first ordinary action of the recovery scene would kick the character straight back into a cascading loop of fainting spell and recovery. Nudging it to 2 fixed the loop. The only way we found that number was by reading players’ stories.

Evaluations still reign supreme. A good death requires human consideration, and, endless fainting loop aside, the failures show up in the prose. The game state looks fine. Our agentic story harness is responsible for using its tools to transmute game state into story, and that system has to be tuned by reading what it actually produces. The things that went wrong in our first rounds were invisible to every metric we had: recovery openings that skipped how much time had passed or why the threat was gone, a rescuer the model invented instead of the companion who was actually standing there, a threat that changed names between scenes. Better evals mean better, faster results.

Make the rules visible, and make them easy to change. All of this constraint would be a problem if it were hidden from the people playing. So in Atlas, our world-building studio, death is a setting that sits right next to the other rules of a narrative world: its physics, its magic, its technology, its social norms. A creator picks no death, normal death, or permadeath, and the choice shows up for players in the world's intro alongside those other rules, so everyone at the table knows what kind of story this is before the first turn. Death: have it your way!

Each of those settings also has to handle its special cases. In a cozy mystery, the protagonist and the core cast might be immune to death, but a murder mystery does typically require a murder! A world built on The Crow needs its hero to die in the first scene and come back as a spirit of vengeance, which breaks every rule above. In practice this is more of a mechanical consistency challenge than a narrative one. Most AI narrators are happy to make exceptions, since saying yes is the problem in the first place. Once you've gotten good at making the system follow the rules, you have to deliberately build in the places where a specific story is allowed to break them, and you have to do it without carving out a new control flow for every world that wants to do something interesting with death.

Conclusion

A narrator who can't hurt you can't tell a story that matters. Death is the far end of consequences. If the story can handle that, the smaller stakes are real too. That's what the story harness is for: a DM that commits to real choices, backed by rules, authored playbooks, and memory, so that losing means something and the system isn't fighting its own tendencies. When done well, it’s a system that uses structures to set up the player for an interesting variety of meaningful choice, where your agency isn’t taken away by capricious consequences but you might still get a little frisson of temptation to do something outrageous if it would be in character.

You stare at the charred, motionless form of your mentor. Smoke curls slowly from the dragon's nostrils and mouth, each tooth as long as your arm, as it considers you with glittering eyes. You have a moment to act: to pick up the next set of cutlery, to flee, perhaps even to reach your mentor's homing medallion and teleport both of you out of here.

What do you do?