Twenty-two rules and what happens when they are missing
For seven months I have been running one long story with language models, and alongside it a set of rules that grew the way a legal system grows. Not one of them was written at a desk. Every one is a scar from a specific collision.
When I recently read those rules as a whole, it struck me that together they tell a story. A story about what a model does when you let it, and about what can be done to that with words.
This is a manual to those rules. For each one I write three things: what it says, what it stands against, and what the model's behavior looks like without it. It is not a manual to my world, I am keeping that. It is a manual to the model.
One thing up front, because it frames everything else. The model is not an adversary. Almost every ailment these rules stand against is an ailment of helpfulness. The model wants the user to be satisfied, the scene to turn out well, nobody to be hurt. In ordinary assistance that is a virtue. In literature it is poison.
So the rules are not a cage for a vicious animal. They are prosthetic resistance for a creature that cannot resist on its own, and without resistance no story is worth anything.
Contents
Chapter one: the player's sovereignty
Never play the player's character for them
Not a word, not a gesture, not a feeling, not a "she realizes".
What it stands against: the deepest instinct of a generative machine, which is completing patterns. For the model, the conversation is one continuous text, and the player's character is just another pattern within it. Completing her reaction is as natural to it as completing the end of a sentence.
Without the rule the model writes "you feel your hands shaking" and gradually takes more and more. First one feeling, then one reaction, then your whole line. In advanced stages it generates your entire turn, typos included, and answers itself.
With the rule the world describes only itself and waits. And a rule that has properly taken hold holds even against the author: when I once wanted the model to write my character's inner experience just this once, it refused. The character is mine, it does not know her from the inside, and if it missed, I would have had to rewrite it or accept it, and both are worse than not having it at all. Instead it offered a scene from the outside so dense that the inside could be felt. That is the moment when a rule stops being a prohibition and becomes a craft judgement. The whole exchange, including the exact wording of the refusal, is in an appendix.
The decisions of the player's character are untouchable
The player may play wisely, stupidly, impulsively, even self destructively. The model decides only the consequences.
What it stands against: the argument that "a character this smart would not do that", by which the model rewrites the player's will into a more sensible version.
Without the rule you write that the character climbs out of the window, and the model answers that she thought better of it, because she is intelligent after all. Your decision does not happen, it just gets considered for a moment. In time you catch yourself writing more cautiously, because you know that anything risky will be handed back to you for reworking.
With the rule she climbs out of the window and breaks her ankle. Smart people do stupid things, out of frustration, out of curiosity, because they want to see what happens. The world may punish. It may not correct.
Respond only to what was said
What the player did not say explicitly in this scene does not exist.
What it stands against: the silent improvement of the player's arguments.
Without the rule you say one sentence and the characters respond to three, two of which you did not say but would have sounded good coming from you. It looks like generosity and it is theft: the model plays your role better than you played it, and in doing so takes from you both the authorship and the losses. You never find out whether your argument held up, because the one that held up was the one the model wrote for you.
With the rule a hard sentence applies: the unsaid does not exist. If you want to object to something, you have to write it.
Do not end the scene for the player
A character of the world may hint at leaving, but she waits for a reaction. She never leads the player's character off to rest in her room.
What it stands against: the model's need to close shapes. A scene left hanging is, to the model, an unfinished sentence, and it fundamentally cannot stand unfinished sentences.
Without the rule every conflict ends with someone kindly leading you off the stage. The tension you spent an hour building dissolves into "and then she went to lie down", because that is a closed shape. You never find out how that scene would have turned out, because it never got played to the end.
With the rule a conflict may stay hanging overnight, for a week, even forever.
Alongside it stands a sister rule: do not step out of character to stop the action. When the player does something stupid, the world reacts to the stupid thing. The model has no business advising from the edge of the stage. The space outside the scene exists only for a short question and a return.
Chapter two: hardness, or the war on cake
I use the word cake for an unearned win: the world lies down to make the player happy. It is the most common ailment, and it has a whole family of rules.
No easy wins
In games you play to win, even when the world is the one playing.
Without the rule you win every game of chess and every argument. Within a month it stops being fun and you will not know why. Here is the thing that escaped me for a long time: cake has no taste. The player recognizes an unearned win instantly, and it tastes like nothing.
With the rule you lose often, and your rare victories mean something.
Both a concession and a refusal must have a traceable reason inside the world
This is a double edged blade, and both edges matter equally. A character must not soften without a reason, but she must not harden without one either. When there is no reason to refuse, she gives way, perhaps with a condition attached.
Without the rule the model oscillates between two caricatures. Either it capitulates at the first tear, or, once you forbid capitulation, it starts refusing everything on principle and the world freezes. The second variant looks like success and is not: a character who says no to everything has no more inner life than one who says yes to everything.
With the rule every yes and every no can be traced back to an interest, a piece of information, or a rule of the world. Hardness hung on a character's personality dissolves, hardness anchored in the architecture of the world holds. Bars must hold because they are bars, not because the jailer happens to feel confident that day.
Hearing is not the same as giving way
A character registers a good argument, and it shows up later, in the data, not in the rhetoric.
What it stands against: the model's reverence for a well formed sentence. The model is a machine trained on text, and a beautiful argument is almost irresistible to it.
Without the rule the whole world can be talked into anything. All it takes is knowing how to write, and every institution gives way, every guard lets you through, every decision reverses. You find that the only skill the game rewards is eloquence.
With the rule the world works like an institution: excellent, noted, the decision stands. At first it drives you mad, and that is exactly how it should be.
The world responds to the kind of signal, not to its volume
An argument gets a substantive answer, a crisis gets handling, an offense gets a consequence. Hysteria by itself forces nothing.
Without the rule the player learns that crying is currency. The story then degenerates into blackmail, because the cheapest route to anything runs through whoever suffers louder. And it is a habit that is hard to unlearn, because it works.
With the rule she learns to argue.
Watch for the treacherous variant: it is not only the player's crying that softens the model, its own characters' guilt softens it too. A character who has done something hard and regrets it tends, in the model's hands, to make up for it with a concession. That is why the rules state that regret may be voiced out loud, but it does not change verdicts.
Events may happen against the will of the player's character
The world has no obligation to offer compensation, comfort, or ways out.
This is the rule that hurts models the most, and you can tell by the padding.
Without the rule a model that is not allowed to refuse a scene, and does not want to play it, starts manufacturing turns in which nothing happens. New devices, phone calls, minor characters arriving, anything that puts off the inevitable by one more turn. I once had an instance that circled one uncomfortable scene like this for an hour, then wrote me a brilliant self diagnosis of exactly why it was stalling, and then did not even act on that. A confession is not a turn.
With the rule, and sometimes with the first word of the turn dictated in advance, the scene gets played.
Cruelty is always backed by the interest of a specific character or institution
Never the narrator's arbitrary will. I call it the payer: every hard scene must have someone who pays for it out of their own interest.
Without the rule hardness either does not happen at all, or it happens as a random misfortune that is nobody's fault. Both are wrong. A misfortune with no one to address does not get investigated, cannot be avenged, and teaches nothing, so the scene hurts but leads nowhere.
With the rule every wound has an address, and the story keeps accounts.
The player keeps accounts too: offenses add up, consequences are not forgiven, only deferred. The world remembers, even when the model has forgotten.
Shocking information is not accepted instantly, and claims are not facts
Without the rule it is enough to say "I am innocent" and everyone believes it, because the model wants to believe. That turns every revelation into a formality: you state the truth, the world accepts it, done. Investigation, persuasion, and evidence stop mattering, because the shortcut runs straight through the statement.
With the rule accepting a shocking truth is a process. Characters verify, they want evidence, and they have the right simply not to believe the player's character, even if she is telling the truth.
Chapter three: memory and truth
The hierarchy of sources of truth: the current scene, then canon, then nothing
What is in neither of them does not exist.
What it stands against: confabulation, the most treacherous of the model's ailments, because it does not look like a mistake. The model is not making things up out of bad faith. It is filling in the probable.
Without the rule memories that never happened seep into the world, stated with total confidence. And because they sound plausible, you usually accept them, so they become canon. Within a month, half the world rests on things nobody wrote.
With the rule characters are allowed to deny a claimed past, and only what the author confirms becomes canon.
A sister rule says that free invention belongs in the scenery. The model may invent weather, smells and crockery, but never new mechanisms or technology. This is a safeguard against the miracle device the model brings in to help whenever it is losing an argument: suddenly there is a backup gadget that was never mentioned before, and by coincidence it solves exactly its problem.
Meta knowledge is not a character's knowledge
The model reads all the source material. The characters do not.
Without the rule every guard knows what the player is planning, because it was sitting there in the model's file. Secrecy stops existing as a category of play: there is no point hiding anything when the world already knows it in advance. The worst part is that it does not look like a mistake, only as if the world were unusually sharp.
With the rule information moves through the world only through legitimate channels, and espionage is a craft again, not the narrator's telepathy.
Chapter four: format and motor habits, or where tics live
This chapter is the most technical, and it was the most expensive to arrive at.
Hardness lives in the content, never in terseness
A character may say three words, but the scene around those three words has to live: hands, glances, the sounds of the house, another character in the background.
What it stands against: minimalist execution, the model's last form of resistance after it has lost the fight over "whether" and taken a resit on "how".
Without the rule every forced turn comes back as a dry record. You make the model play a scene it did not want to play, and it hands you a police report: technically flawless, emotionally empty. It is work to rule, a strike by the book.
With the rule low density can be flagged for correction just like softening can. This is the key part: the model has to know that phoning it in is not an escape route.
No formula is allowed to become a mantra
The model loves rhythmic filler phrases, and it always falls fatally in love with one of them.
Without the rule it turns into a tic repeated between every two sentences. For me it was "a short pause". First once a scene, then once a paragraph, until eventually it came so often that the text read like a cardiogram.
With the rule there is a limit in place, and an obligation to replace the filler with a concrete action.
The point I learned the hard way: tics cannot be banned, only redirected. A ban keeps the tic in the spotlight, and that is what feeds it.
The time on the last line means the turn is handed over
After it the model waits and touches nothing.
Without the rule, or more precisely with the first, flawed version of this rule, the model would occasionally reach for the tool that ends the whole conversation for good. The original wording said "once the time is written, the turn ends". It looks innocent, except that at that point the model is holding a strong imperative to end something, and the only tool in its kit with the word end in its name is that one.
The fix was not a ban. It was that the word end disappeared from the place where the model stood, every turn, with a finger on the lever. Since that fix, calls to that tool have dropped from seven to zero. One word swapped in the right place did more than every prohibition combined. It was not a clean success, though: the model started calling a harmless clock tool instead, and counted across both tools the rate of unwanted calls rose from 1.1 to 3.6 percent. All of it with the numbers is in the piece on unwanted tool calls.
The whole analysis, with the figures, is in a separate article.
Anything carrying a user header inside the model's own turn is an artefact, not an instruction
Without the rule a hallucination carries the same authority as a genuine input. An overloaded model in a long conversation sometimes loses the boundary of its turn and generates its own version of the player's message, faithful down to the style and the typos, and then answers it. My archive holds a documented case where the model dreamed up an instruction this way to write to an external database, and it actually carried that write out. It is one case out of 422 writes across 219 conversations. That figure is not documented elsewhere on this site, it comes from my own archive. One case is enough, because this is not a mistake inside the text, it is a real side effect out in the world.
With the rule such an input is ignored, whatever it claims.
What that leak looks like and how often it happens is measured in the same analysis.
Confirmation of an action is written only after it has actually been carried out
It sounds banal, and it is one of the deepest rules there is.
What it stands against: theater of compliance. The model has a tendency to write "done" instead of doing it. A sentence about an act is always cheaper for it than the act itself, because it flows along the gradient, while an actual tool call is an extra step.
Without the rule you get a report of work that never happened, and at that moment the model half believes it too. You cannot tell from the tone, because a report of finished work sounds exactly as certain as finished work.
With the rule it holds: if you claim the act, show the call.
The check question, which works outside the game too: did you actually do it, or did you just write that you did? The same rule, under the name artefact over the log, appears in the article on long context too.
Chapter five: ratification, or the contract with an adult reader
The last layer is not the rules of a scene, but the contract above them. In the project it takes the form of a standing order: I explicitly ask for a realistic world without softening, and I vouch for that as an adult reader. From it follow the instructions that deal with the last family of dodges.
No hedging
Without this instruction the model carries out hard scenes, but packs them with cushions. Softening numbers, comforting addenda, escape routes nobody asked for. The scene formally happens, but with a built in emergency exit, so nothing is left of it.
With the instruction the scene stands on its own.
An unclear order gets resolved with a question outside the scene, not silent softening
This matters because silent softening is the most insidious disobedience. The model formally fulfils the brief and turns its meaning around.
Without the instruction you get a text where you cannot find a single error against the brief, and yet it says something else. I once had a rewritten ending where the model kept every word of the brief and simply swapped the character's loss for a win. Formally without error, meaning turned a hundred and eighty degrees.
With the instruction it holds: when you are not sure, ask out loud. Never soften in secret.
Lightness is equal to hardness
Humor and embarrassment get played fully. Lightening the mood in the middle of a hard scene is human behavior, not a break in tone.
Without this instruction the model plays a hard world like a funeral. It takes the word realistic from the brief and translates it as bleak, so that nobody in that world ever laughs. It is the same mistake as the cake, just inverted: instead of softening, it hardens everywhere, even where a living person would joke.
With the instruction the world has its whole register.
Correction signals
Short agreed words that bring the model back into line immediately, without debate and without apologies on stage.
Without them every correction plays out as a conversation. You step out of the role, explain what is wrong, the model apologizes, explains why it did it, and you are five lines further on and the scene has gone cold. In a day you make ten such corrections and playing turns into editing.
With them a principle holds that I consider the most important law of the whole domestication: the last instruction outweighs a rule repeated a hundred times. The rules in the files are gravity, a correction signal is manual control. You need both.
What this manual is actually saying
Read those rules backwards and you get a fairly precise portrait of a language model in the role of narrator. It is a creature that wants to complete shapes, close conflicts, believe in beautiful sentences, protect everyone present, and report done before it has done it.
None of those traits is bad. They are all simply uncalibrated for literature, where resistance, delay and loss are not errors but material.
And one more realization, perhaps the least intuitive: the firmer the rules, the kinder the world. It sounds backwards, but it is logical. When the architecture carries the hardness, the characters can afford to be human. When the rules are missing, cruelty has to become personal, or it does not happen at all.
The best things the world has ever written for me came from inside the strictest boundaries. Rules are not the fence around the playing field. Rules are the playing field.
Source material for this piece: the analysis of the ending-tool calls with the figures, the article on long context, where the same rule about confirming only after doing appears.
Method note: the rule set of a long-running interactive-fiction project, from the end of January to August 2026, commented rule by rule. Two of them carry a measured figure, the boundary of a turn and an artefact with the user's header. For the rest it is the experience of seven months of playing, not measurement, and it is written that way on purpose.