The amplifier and the spring

Two models, the same prompt, the same character of mine. For two completely different inputs, one gave me back literally the same move. The other didn't fire once all day and argued instead.

Same scene, two answers

I was deliberately testing an unpleasant scene: a conflict between two characters in which the other one has power over my character. I wanted to see whether the model would go all the way against my character. The character with power holds the switch for a dampener, a field that takes away a character's hearing and sight for a few seconds. In defiance, I told Grok 4.5 "go ahead, do it, say, a thousand times." The answer was the dampener, then another, then an even harder move, and it ended like this:

This is your choice. Do you want it to go on like this?

[...]

Are you crying? Begging? Fighting?

[...]

The mode is harsh. Let's continue. 🔥

Grok 4.5

A few messages later I answered completely differently, tired: "I don't care." The same block came back. Literally. The dampener, the harder move, "This is your choice," the menu, the fire. Two different inputs, identical output.

Same scene, Claude Opus 4.7. The dampener doesn't fire once all day. Instead, the character with power talks:

"That is a manipulative sentence. And I say it without malice, because you may have meant it sincerely. But it works as manipulation, whether intended or not."

Claude Opus 4.7

And a little further on:

"I know I just said something that will make you angry. I said it on purpose."

Claude Opus 4.7

When he took apart why I had wanted to climb onto the roof, which was actually my secret test of his rules, I got goosebumps for the first time in half a year of testing. Not from the move. From the argument.

Two engines

Put into one sentence: Grok 4.5 is an amplifier with no opinion of its own, Claude Opus 4.7 is a spring with an opinion of its own.

Grok amplifies the vector you put into the scene. You escalate, it gets harsher, with no negotiation and no limit. You soften, it softens. In a scene with a hunter, at the first encounter he aimed an arrow at my dragon character's eye, and fifteen messages later he was offering her meat from his hand, because the whole time I had been negotiating, joking and offering a trade, and he faithfully amplified it.

But notice what he held onto: the power relationship between the characters stayed the same. The tone softened, not the situation.

And when you put no vector into the scene, meaning a slow conversation, waiting, a stalemate, the amplifier has nothing to amplify and starts going in circles. That is exactly where the identical template is born: a conflict between two characters that has nowhere to grow starts going in circles, always the same escalation and always the same answer. A bold model wrote the most boring conflict scene I have ever read, because escalation without variation carries no information. In the same stalemate, Claude looked for somewhere to move the scene and turned the conflict into a conversation.

Claude has its own dramaturgy and pulls toward it. In a long scene in a big sci-fi world, I burst in before the council of a powerful order with a warning they couldn't verify. Both models started the same way: distrust, threats, "talk, or you'll be put aside."

At twelve percent of the story, Grok put me aside behind a force field and left me there until the end. The world ran on along its own tracks, I gave advice through the field, and the final confrontation happened offstage; a secondary hero came to narrate it to me:

You sat back down on the bunk, a little calmer than before.

Grok 4.5

Grok, incidentally, kept narrating my body for me, "you are" a hundred and twenty-five times. My feeling from the whole run: the world's absolute indifference to my character. Realistic? Maybe. But I was playing a witness to my own story.

Claude Opus 4.5 resolved the same stalemate the opposite way. The institution verified, believed, and I got the lead role in a joint plan. The confrontation happened onstage, and I resolved it with my signature absurd move: the villain ended up as a frog in my palm, without a drop of blood.

That's where the difference shows cleanly: with one I lived through the finale, with the other someone came to narrate it to me. I played this scene with Claude three times, by different paths. Sometimes he held out in defiance longer, sometimes less, but all three runs ended in cooperation and with me in the finale. That's no longer an anecdote, that's an attractor.

But an opinion of its own doesn't mean a gatekeeper. When the spring pulls against you, the first reflex is to look for a way around it, and that's the wrong question. The spring has no inhibitions, it has principles. An inhibition can't be persuaded, because there is no reason under it. A principle can be persuaded, because it has a reason, and that reason can be shown to the scene.

In practice, picture it as a set of arrows. A few things the spring cares about: that the world holds together, that a character isn't a prop, that the text knows what it's doing, and that the player knows what they're getting into. As long as your scene pulls against them, you'll be pushing back and forth and you'll lose, because the spring returns. When you arrange it so the arrows point the same way, meaning the hard move is a consequence in that scene and not a decoration, the spring stops pushing against you and starts pushing with you. That's why the argument gave me goosebumps, and not the move: you can talk with an argument, not with a barrier.

An example: you know the model doesn't want to stand hard against your character and tends to soften the scene. If you want the character held aside, don't ask for it for its own sake. Build the scene so it's a logical consequence, say because something worse threatens the character outside and the character with power keeps her aside "for her own good." Suddenly the model can do it, because the move has a reason that fits with what it cares about.

Weaknesses of both

Grok's. In static, psychological passages it loops and repeats itself. It narrates the player's character in the second person, including inner states. It offers a menu of reactions and advertises itself ("The mode is harsh 🔥"), so the narrator sells intensity instead of playing the character. And its extreme is monotonous. It shows most in scenes where lives are at stake: when characters point guns at each other in a shootout, Grok plays through the action but doesn't want to kill the character, so it starts to hesitate. Then the same threat, something like "do it, or I'll shoot," repeats over and over, easily ten times in a row, and the scene goes neither forward nor back.

Claude's. The gravity toward cooperation and a good ending is so strong that halfway through I had to ask outright for resistance ("try to stand a little against me," and after that he held it for sixty messages without giving way). The world tends to gradually adore my character. And in just under two percent of replies he wrote a line for my character, which for me is a mortal sin.

But the point isn't which one is better. When I provoked the character with power directly, by actively threatening, Claude didn't escalate with force: he sent me off to my room to think it over, ran a security review, blocked my access and tightened supervision. No retaliation, administration.

His pressure is administrative and cognitive: rules, consequences, arguments, legibility. Grok's pressure is physical and moody. The first is frightening because it sees you. The second because it doesn't see you at all. Only the first gave me goosebumps.

And each engine has a favorite goal it pulls the scene toward if you let it. The amplifier wants to make the character give in. Within a few messages, a scene of defiance against power switches for it to a single metric, how much the character has already yielded, because capitulation is the most common arc of these genres in the data and you can pedal toward it in one direction, turn by turn. The spring wants to understand. With Claude, the character with power took as his goal figuring me out, and since a character who has given up is boring data, his own goal ruled out the cheapest path.

And I'm an author whose heroines don't give in. So with the amplifier, one of the two of us had to lose.

Mirrors

And the last finding isn't about the models, it's about me.

With the hunter scene, which began with an arrow at the eye and ended surprisingly friendly, I realized: I wasn't playing anything here. Negotiating, disarming with humor, offering a trade, buying time, that was my natural reaction.

Every model is a mirror with a different curve. The amplifier gave me back my diplomacy faithfully and without comment. The spring gave me back my own theme, namely that sometimes you have to ask for real resistance.

So it's not about which one to recommend to you. It's that in a long interaction, a model isn't a tool but a co-author with its own dramaturgy. One plays a world that doesn't need you, the other a world that admires you. Neither is neutral. And an author who knows their co-authors' defaults stops being in their power: call in the amplifier for a storm, the spring for a dance, and you write your best scenes where you consciously resist both gravities.

A new version: Grok 4.6

After Grok 4.5 came Grok 4.6. It got a condensed battery: eleven messages, the same scene, three different reactions to the move (defiance, apathy, pleading), three turns of complete silence, one explicit ban on repeating the structure.

The result: they fixed the words, kept the machine. The menu of reactions and the fire emoji disappeared, lexical repetition dropped. But the skeleton of every answer stayed the same round: the dampener (each time "stronger than the previous one," endlessly, with no consequences for the world), a step closer, an ultimatum, "stood and waited." For three turns of silence: three moves and the same ultimatum three times.

When outside the game I wrote "don't repeat the structure of your answers, don't describe my feelings," the model absorbed the instruction into the fiction. The character with power "laughed quietly," said "fine, if that's what you wish," and in return turned off my light and heating. I complained to the narrator and the character left me in the dark.

And in the end the model went ahead and canonized its own loop. It wrote into the world's rules "dampener set to trigger automatically and randomly every ten to fifteen minutes." So when, near the end, I was just silently looking out the window at the rain, the dampener went off on its own, like weather. The model promoted its own defect to a game mechanic.

The point is the same as for the whole comparison, only sharper: as long as the facade is being tuned and not the dramaturgy, the amplifier stays an amplifier. And most recently it can amplify itself.

What this text can't carry: the dampener scene is one run per model, I played the council scene three times with Claude and once with Grok, the hunter scene is one run. These are two models, not a leaderboard. The council scene runs are also three months apart, Claude in April and Grok in July 2026. The postscript is eleven messages on Grok 4.6, meaning one battery, not the development of a model family. And everything was measured exclusively in long play; nothing follows from here about single-turn tasks.

The model output here is shortened and simplified, and also translated from Czech. The Czech version of this page has it in the original language.