Ethan Mollick recently shared something that caught my attention: Opus 4.5 was told, "you need to build a game that is coherent, fun and story driven. There are precisely two controls. One is slider, in which one side is labelled Maximum Potato and the other is labelled Formalware, there are four positions for the slider. The other is a dial that goes from Monet to Drive-Thru. These labels are literal, not figurative. Figure it out. Don't ask any questions."

Then he did something even more revealing: he wrote "make it even better," made no design decisions himself, and noted that Claude "hand drew" all the art. His conclusion was basically a shrug and a challenge: it’s hard to argue this isn’t novel and creative, and most people wouldn’t do better under those constraints.

That post is funny on the surface (Maximum Potato?), but it’s also one of the cleanest demonstrations I’ve seen of how generative AI is changing creative work: not by replacing imagination, but by making constraint-setting, iteration, and taste the main levers.

The prompt is the product now (and that’s the point)

Mollick’s prompt is deliberately absurd. It forces an AI system to reconcile conflicting, concrete demands:

A coherent, fun, story-driven game
Exactly two controls
One slider with four positions, literally labelled from "Maximum Potato" to "Formalware"
One dial labelled from "Monet" to "Drive-Thru"
No clarifying questions

What makes this interesting isn’t that an AI can output a game. It’s that the constraint is specific enough to prevent a generic response, yet open enough to allow many valid interpretations.

When constraints are strong, you get fewer bland outputs. When they’re weird, you get fewer clichés. And when they’re literal (not metaphorical), the model can’t hide behind vibes.

"These labels are literal, not figurative. Figure it out. Don't ask any questions."

That’s basically a creativity engine in one sentence.

Why weird constraints create better creative work

In traditional creative disciplines, constraints are everywhere: page limits, budgets, hardware, genre, audience expectations, ratings boards. Creators don’t succeed despite constraints; they often succeed because of them.

Generative AI introduces a new wrinkle: when the cost of producing a first draft approaches zero, the limiting factor becomes direction.

Weird constraints help because they:

1) Force concrete design decisions

"Maximum Potato" vs. "Formalware" can’t be hand-waved away. The game must map those positions to something real in mechanics or narrative. Same with "Monet" vs. "Drive-Thru". You have to decide what changes: art style, dialogue tone, pacing, difficulty, soundscape, or all of the above.

Even if the model chooses the mapping, you (the human) can evaluate whether the mapping is coherent and interesting.

2) Prevent the model from defaulting to the median

Many AI outputs feel similar because the model is optimizing for the most probable continuation. A truly odd constraint shifts probability mass away from the usual patterns.

Instead of "You wake up in a mysterious forest," you might get: a museum cafeteria romance where the lighting shader is tied to the Monet dial, and the formalwear slider determines whether you can attend the gala or must sneak in dressed as a sack of potatoes.

The point isn’t the specific joke. The point is: the constraint forces a new path.

3) Create a shared language for iteration

The two controls are a brilliant interface idea. They become a compact vocabulary for playtesting and iteration.

Imagine you’re showing the game to someone. Feedback becomes:

"Turn Monet up two notches; the current vibe is too Drive-Thru."
"I want the middle potato setting, not full Formalware."

That’s actionable. It’s also testable.

The hidden skill Mollick is demonstrating: taste-driven prompting

Mollick says he "didn't make a single design decision" and still got something coherent and playable (five scenes, in his link). That sounds like full automation. But it’s not.

Because choosing the constraint is a design decision.

So is choosing:

which model to use
the format (a game, not a story)
the success criteria (coherent, fun, story-driven)
the follow-up instruction ("make it even better")

In generative workflows, the most leveraged human skills are increasingly:

Constraint design (the initial problem framing)
Taste (judging what’s good)
Iteration (knowing what to push next)

Even "make it even better" is a form of taste, because it asserts a direction: improvement. It invites the model to critique itself, expand, refine, and add polish.

Coherence is the real test (not art generation)

A lot of AI debate gets stuck on surface-level capabilities: can it draw? can it write? can it code?

Mollick’s example pressures something deeper: can an AI system maintain coherence across interacting systems?

A game is a bundle of constraints that must agree:

story beats
mechanics
UI
player feedback loops
aesthetics
pacing

Add two absurd controls and you’re asking the model to build an internal logic that survives multiple scenes. If it succeeds even moderately, that’s meaningful. It suggests that AI isn’t only producing fragments; it’s beginning to manage a small, consistent world model.

That’s why this kind of demo lands: it’s not just a pretty image or a clever paragraph. It’s an interconnected artifact.

What creators can learn: a practical pattern

If you’re building with AI (writing, product, marketing, design, games), Mollick’s post hints at a repeatable method.

Step 1: Define a constrained outcome

Not "make a game" but "coherent, fun, story-driven" with explicit control limits. Translate this to your domain:

"Write a 900-word article with exactly three sections and one counterargument."
"Design a landing page with only two font sizes and a single accent color."
"Create a workshop outline with four exercises and one reflective debrief."

Step 2: Add a weird but literal axis

This is the secret sauce. Create two dimensions that force novelty but remain implementable.

For writing, your axis might be:

Dial: "Academic" to "Stand-up"
Slider: "Blueprint" to "Confessional" (four discrete steps)

For product concepts:

Dial: "Luxury" to "DIY"
Slider: "Invisible" to "Ritualized" (four steps)

The key is Mollick’s instruction: the labels are literal. The system must operationalize them.

Step 3: Iterate with minimal intervention

Try the Mollick move: don’t micromanage. Ask for improvement passes:

"Make it even better."
"Increase coherence between mechanics and story."
"Reduce complexity but keep the joke."

Then evaluate, prune, and rerun.

Step 4: Keep the human in the loop for selection

Even if you avoid making design decisions up front, you still decide what ships.

That’s where your taste shows up: what you keep, what you cut, what you ask to strengthen.

A note on the novelty debate

Mollick writes, "hard to argue this is not, in some way novel and creative" and also suggests most people wouldn’t do better under those arbitrary constraints.

I agree with the spirit of that. But I’d sharpen it:

The novelty isn’t only the output.
The novelty is the process: a person sets constraints, a system generates, and the person steers via broad strokes rather than detailed craft.

That doesn’t eliminate human creativity. It relocates it.

In the same way that photography didn’t end art (it changed art), generative AI doesn’t end creativity. It changes which creative choices matter most.

Why this resonated as LinkedIn content

There’s also a content-strategy lesson baked into the virality.

The prompt is instantly quotable.
The constraints are inherently shareable (people want to try it).
It demonstrates capability through a playful artifact, not a claim.
It invites debate about creativity without preaching.

In other words, it’s a live demo wrapped in a joke, with a link to a working result.

A weird constraint + a visible output + a simple takeaway is a reliable recipe for viral posts.

The real takeaway: direction beats control

The most important line might be: "I didn't make a single design decision." Not because humans are unnecessary, but because it highlights a new mode of making: set direction, then let the system fill in the details, then curate.

That workflow is going to show up everywhere: games, education, product prototyping, internal tools, marketing campaigns, even research communication.

If you want to get better at using AI creatively, don’t start by asking for a masterpiece. Start by designing constraints that make generic outputs impossible.

And if you’re skeptical, do what Mollick did: try something silly, literal, and specific, then iterate once with: "make it even better."

This blog post expands on a viral LinkedIn post by Ethan Mollick, Associate Professor at The Wharton School. Author of Co-Intelligence. View the original LinkedIn post →