Text-based AI RPGs often hit the same wall: whatever the hero declares, the narrator obediently describes as a success. Fooled the guard, jumped the chasm, persuaded the king – everything works on the first try because the model defaults to agreeing with the player's request. A story without rejection quickly loses its flavor: there are no stakes, nothing to plan, and reading about yourself as an invincible hero gets boring. Let's break down why this happens and how a single twenty-sided die brings real uncertainty back to the scene – the very reason people sit down for skill checks via d20.

The Problem of Agreeableness: Why AI Never Says "No"

Illustration: a silhouette of a player at a keyboard looking at a screen where an AI character nods and smiles at every request. Around the screen are text bubbles saying 'Yes', 'Of course!', 'Whatever you say' – a feeling of infinite agreement

A language model is built simply: it continues the text in the direction you set. Write "I convince the guard to let me pass" – and the model describes how the guard hesitates, sighs, and steps aside. Write "I chop off the dragon's head with one blow" – and the dragon falls. The model does not evaluate the hero's chances; it builds the scene to fit your prompt. A refusal in its worldview is a rare, unlikely branch of continuation, while agreement is the standard.

In the short term, this even feels good: your hero is agile, witty, and invincible. The problems start around the third or fourth scene. If every planned action comes true, there is nothing left to plan. The choice between "bribe", "intimidate", and "sneak past" loses weight: any option will work just because you chose it. The story stops being a story and turns into a dictation – you dictate events, and the AI formats them with beautiful prose.

In tabletop RPGs, this doesn't happen because there are two forces standing between the player and the outcome: a game master who can say "that won't work," and a dice roll that can result in failure even with a solid plan. These two create resistance. Because of them, the prompt "I convince the guard" has stakes: what if it doesn't work, what if he calls for backup, what if they remember your face? This resistance is exactly what makes a scene a true scene, rather than a retelling of your wishes.

In AI role-playing games without a built-in check mechanic, there is no resistance. The default model is on the player's side – not out of politeness, but by design. To bring stakes back, you either have to ask the model to reject itself (which works worse than it sounds), or give it an external arbiter to obey. The easiest way is the number rolled on a die.

What is Narrative Tension and Why Stories Need It

Narrative tension is the gap between "what the hero planned" and "what actually happens." As long as that gap exists, the reader or player leans forward: they care how it ends. As soon as the gap disappears and everything planned happens in a snap, the same scene turns into a logbook: this happened, then that, then another thing.

Anton Chekhov liked to fix this with a single detail. The gun on the wall in the first act works not because it looks nice, but because it promises a shot – and we spend the whole play not knowing at whom. Alfred Hitchcock explained suspense through the same mechanic: a bomb under a table is terrifying right up until the explosion, not after. If the hero knows in advance they will defuse it, there is nothing to watch.

In tabletop RPGs, this gap is maintained by the dice. The game master doesn't pull the hero's luck out of thin air or distribute it fairly – they leave it to chance. That's why people at the table laugh when a rogue with ten successful lockpicks fails the eleventh on the simplest lock: a scene that should have taken one sentence suddenly becomes the story of the evening.

Hence the main paradigm shift, which can feel unusual to accept: the hero's failure is not the breaking of the story, but its fuel. A successful deception closes a scene; a failed one opens three new ones – the guard called for help, the plan is ruined, the hero must improvise. Plots are driven not by victories, but by complications. Any screenwriting textbook will say the same thing in different words: without an obstacle, there is no arc; without a price for mistakes, there is no growth. The storyteller's job is not to protect the hero, but to honestly maintain the unknown until the very roll.

How a d20 Brings Uncertainty Back to Text RPGs

Close-up of a d20 flying over a game board with a map and character sheet. The numbers on the faces are blurred from spinning – the moment of the roll, the outcome is unknown. The tension of uncertainty.

The die solves one problem: it takes the right to say "no" away from the storyteller and gives it to a number. As long as the refusal comes from the scene's author, it sounds like subjective bias – "why didn't it work right now, I had a good idea." When the refusal comes from rolling a four on a twenty-sided die, there is nothing to be offended by. A four is equally indifferent to both bad ideas and good ones.

From there, the storyteller keeps three gradations in mind – and determines each one themselves, based on the scene and the hero.

The most important thing is what happens after a failed roll. The story doesn't roll back to the "try again" crossroad. Time has passed, the thief messed with the lock too long – a patrol is already around the corner. The servant they tried to bribe got suspicious and will report to their master. The previous method has closed, forcing other paths to open wider. Repeating the same action costs more: the difficulty increases, and you don't get the same scene twice.

The extremes of the scale operate independently of difficulty. A natural 1 is a critical failure: the hero pays a price right in the scene – a wound, a broken blade, revealed intentions, an alarm raised. A natural 20 is a critical success: the result echoes further down the plot with an advantage, an ally, or a clue. You can read more about the check mechanics in this example of d20 in AI games – showing how difficulty is calculated and why it isn't fixed.

Prompts and Tricks: How to Ask AI to Disagree with You

Screenshot-style illustration: two versions of dialogue with AI – on the left a bad prompt and the response 'Of course, everything worked out!', on the right an improved prompt explicitly stating action difficulty and a response with a roll and failure.

If you play through a regular chat with an AI model, its agreeableness can be treated with a prompt. Not completely and not forever, but noticeably. Let's break down exactly what to write in the system instruction so the narrator stops nodding along to your every move.

The first and main thing is an explicit ban on automatic agreement. A direct phrasing works better than hints. An example system prompt:

Compare two approaches. Before: "I approach the guard and convince him to let me pass." A model without instructions will write a beautiful scene where the guard nods and steps aside. After: The same prompt – the narrator replies: "Difficulty 14, the guard is sober and on duty. Roll: 9. He frowns and tells you to back off." The story has moved, and not where you planned – which is exactly why people play RPGs.

It's also useful to add a list of gradations: "describe mundane actions without a roll," "reject impossible actions in this scene with an explanation." Otherwise, the model will start rolling dice even for "I walk down the hallway."

The limitations are honest. The chat context is finite: after fifty messages, the model forgets its own rules and slides back into "you are doing wonderfully." You have to rebuild the prompt every new session, and long stories fall to pieces. For more on building instructions, see our guide to AI RPG prompts; if you're digging deeper into how a neural network runs a text RPG, other ailments of pure chat are also broken down there.

The prompt approach acts as a crutch. It treats the symptom but doesn't change the nature of the model: it still wants to please you and will slowly drift back toward agreement.

Free Mode vs. Dice Mode: Which to Choose

Two approaches to the exact same story yield very different experiences – and neither is "correct." It's a question of why you are sitting down to play.

Free Mode is for those who came to live through a story, not test their hero's endurance. You plan it – it happens. Fooled a guard, seduced a princess, picked a lock. The narrator describes exactly how, and moves on. This works great when you want romance without misfires, a cozy adventure, or just to roleplay a scene that's been spinning in your head for a long time. The outcome here is determined by the player's intent, and the narrator helps realize it beautifully.

Dice Mode works differently: the outcome of every risky action is decided by a d20 roll, and the difficulty is assigned by the narrator based on the scene. The rogue hero reaches into a pocket – the check is easy. The same hero tries to argue with an archmage – the difficulty spikes. On a natural 1, a failure with a real price occurs: a wound, a lost item, a raised alarm, ruined relationships. On a natural 20 – a success with a bonus that pays off later in the plot. Repeating the same action costs more: for a repeat attempt, difficulty increases by 3 (capped at +6), and the previous method is closed off.

The key point is that failure doesn't rewind the story; it drives it. The guard got suspicious, time was wasted, the plan collapsed, you need a new one. This is where that narrative tension comes from, making the hero's successes actually mean something.

In text RPGs with d20 and an AI Game Master, both modes are usually open to everyone, including free players: each chooses according to their own taste. If you want to understand deeper how checks work and where difficulty comes from, check out our breakdown of skill checks in AI text RPGs. Choosing is simple: if you want a guaranteed story, pick Free Mode; if you want stakes, pick the dice.

How Hero Failure Makes the Story Better: Examples from Games and Literature

A fantasy hero fallen on one knee in a dark dungeon – sword knocked away, torch dying out, path blocked by a cave-in. But behind him, an unexpected corridor with light in the distance has opened. Visual metaphor: failure opens new paths.

If the hero succeeds at everything, there is nothing to tell about them. Frodo would have reached Mount Doom in three chapters, Raskolnikov would have calmly drunk away the old woman's money, and Geralt would have slain Vilgefortz on the first strike. The stories we reread hold up on the moments where the character failed.

Tabletop games understood this before anyone else. In D&D, players still recount epic fails rather than cool successes: how the rogue failed stealth and fell from the rafters right into the throne room, or how the paladin rolled a one on persuasion and insulted the queen. Blades in the Dark went further and wove failure right into the fabric of the game – actions have three outcomes there, and "success with a cost" is often more interesting than a pure victory: you succeeded, but the guards are coming, or you're wounded, or your friend now owes a debt to your enemy.

Video games step on the same pedal. Dark Souls relies on you dying – and each death adds a personal narrative to the world map that you wouldn't have had with an easy victory. Disco Elysium turned failed checks into a source of the best scenes: a detective miserably failing "Composure" in front of a corpse is remembered far more strongly than any successful interrogation.

Literature works the exact same way. Gatsby does not win Daisy. Boromir cannot resist the Ring. These failures ARE the plot – without them, the books would dissolve into descriptions and landscapes. A hero who stumbles, grows; a hero who gets everything handed to them, stands still.

This is why neural networks for text RPGs need a mechanism for refusal. Without it, the AI returns an endless string of successes – and that is exactly why those stories are forgotten by the next evening.

Why does AI always agree with the player – is it a bug or a feature?

It is a feature of how language models are trained: they were trained to be helpful assistants, and a helpful assistant fulfills a request rather than arguing with it. For a dialogue about recipes or code, this is a blessing; for a role-playing game, it is a problem. Declare "I deceive the guard" – and the model joyfully describes a successful deception because it was asked to. The technical nature is the same, but its value depends on the task.

Can I force ChatGPT or Claude to refuse me in a text RPG?

Partially. You use a prompt to set NPC personalities, mortality levels, and the principle "if the chance is less than half, describe a failure." It works unevenly: after a dozen turns, the model forgets the instructions and goes back to giving the hero whatever they want. A more stable solution is an external randomizer that the model doesn't control. Read more about prompt approaches in our AI RPG prompt guide.

What is a d20 roll in an AI game and how does it work?

A d20 is a twenty-sided die from tabletop RPGs. The hero declares an action with risk, the game master assigns a difficulty (a number from 1 to 20), and the server rolls the die. A roll equal to or higher than the difficulty is a success; lower is a failure. A natural 1 gives a critical failure with tangible consequences, and a natural 20 is a critical success. The model doesn't pick the numbers, so it can't play favorites.

If my hero fails a roll, does the story continue or is it game over?

It almost always continues. A failure turns into a plot twist: the guard noticed, the lock broke in your hands, the NPC got insulted and walked away. The scene moves forward with new variables, often more interestingly than the original plan. Hero death is a rare outcome, usually only after a critical failure in a scene where the stakes were already lethal.

Who is the no-dice mode for, and who should use the dice?

No dice is for when you want to guide the hero along an intended arc, experience a romance plot, or enjoy a cozy adventure without the risk of derailing it. Dice mode is for when you want the stakes to be real and victories to feel earned. Both options are equally valid; it's a matter of gameplay taste, not the player's "hardcore" level.