AI makes execution cheap across every creative domain. The bottleneck that remains is making intention legible.
Sometime in the mid-1920s, in a rehearsal hall in Germany, Rudolf von Laban watched a dancer perform a piece he had choreographed. The performance was exactly right. It was also, he knew, already disappearing. When the dancer left the room, the piece would survive only in her body and his memory. If she forgot a passage, there was no score to consult. If a company in another city wanted to perform it, Laban would have to travel there and teach it in person, limb by limb, correction by correction. Music had notation. Architecture had drawings. Dance had only the body that danced it.
Laban spent years building what dance lacked. By 1928 he had published a full notation system — symbols for every limb, direction, duration, and quality of motion, laid out on a vertical staff that could capture a complete choreographic work on paper. It was not a better way to dance. It was a way to say what the dance was — to give choreographic intention a form that could survive the room it was created in, travel without the choreographer, be reproduced by dancers who had never seen the original performance.
What Laban solved was not an execution problem. Dancers had always been able to dance. He solved a specification problem: the gap between what the creator intended and what could be communicated precisely enough for someone else to realize it. Before his notation, that gap was total. After it, choreography could be written down, transmitted, preserved, and built upon.
Other creative traditions solved the same problem in their own time — the musical score, the architectural drawing, the programming language. Each was a different answer to the same question: how do you give intention a form that someone else can execute? The breakthrough, in every case, was not better execution. It was a better representation of what the work should be.
I. When Execution Becomes Cheap
Once intention could be externalized, execution could be separated from the creator — distributed, standardized, eventually commoditized. The composer no longer needed to sit in the orchestra. The architect no longer needed to stand on the scaffolding. Value migrated from the person who could do the thing to the person who could describe what the thing should be.
AI is now doing this across every creative domain at once.
In 2024, the first short films made with AI video generation arrived. The Toronto collective Shy Kids generated hundreds of clips for a single project and used a handful. Director Paul Trillo, producing a music video for Washed Out, generated roughly seven hundred clips and selected fifty-five — about eight percent. The other ninety-two percent were technically competent and creatively useless, because they did not match the intention behind the project. Both teams discovered that standard creative language — “pan right,” “make it more cinematic” — produced wildly inconsistent results. The real work was not in generating footage. It was in finding the precise descriptive language that would make the model converge on what the director actually saw in their head.
The same pattern shows up wherever AI generation has arrived. Suno, one of the leading AI music platforms, generates millions of songs per day — its documentation describes prompts as “vibes, not instructions.” In a 2025 game jam run by Pieter Levels — a competition to build games under time pressure — over a thousand AI-assisted entries were submitted. The winners were distinguished not by the quality of their generated code but by the clarity of their game design. In every case the model was available to everyone. The description was not.
This is not merely an old problem revealed. When execution was expensive, cost itself forced clarity — you had to decide what you wanted before you could afford to make it. Remove the cost, and nothing forces the decision. The question of what to make, once settled by scarcity, hangs open. Execution has become cheap. The constraint that remains is the ability to say, precisely enough for a machine to act on it, what you actually want.
II. The Slot Machine Problem
When execution is cheap and the direction is vague, creation degenerates into a recognizable loop: generate, look, discard, generate again. The filmmaker pulls the lever, watches what comes out, decides it is not quite right, pulls again. The musician types a prompt, listens, adjusts a word, regenerates. The designer scrolls through a grid of variations hoping something will land.
The industry has a word for what this loop produces at scale: slop.
Runway, one of the leading AI video tools, tells the story through its own product evolution. Its earlier models were widely described by users as a slot machine — capable of stunning individual frames, nearly impossible to steer toward a coherent vision. The next version introduced what the company called Director Mode: storyboard tools, camera path controls, character consistency across shots. The upgrade was not primarily about better generation. It was about giving creators more surface area for describing what they wanted. The company discovered, through the behavior of its own users, that the bottleneck had never been the model. It was the interface between human intention and machine execution.
Grimes — who had been among the most visible musicians endorsing AI-generated music — grew publicly disillusioned by 2025. The tools could generate music that sounded like music. What they could not do, without extensive guidance, was generate music that sounded like hers. Without a way to encode what made something distinctively Grimes — the particular collision of genres, the specific textures, the line between polish and rawness — the models converged on the statistical center. Competent. Generic. Interchangeable.
Architecture named the same problem from a different angle. In a 2025 essay for ArchDaily titled “This Is Not Architecture,” Eduardo Souza argued that most AI-generated renders “prioritize visual novelty over design logic.” The images looked like buildings. They were not buildings — they encoded no structure an engineer could build from, no intent a client could revise. They were pictures of a mood, not specifications of a structure. Even firms like Zaha Hadid Architects, long reliant on custom computational design tools, found that general-purpose AI could not bridge the gap between what a render looks like and what a building needs to be.
Slop, in all of these cases, is not only a failure of the model. It is a failure of specification. The machine can generate anything. It does not know which anything you meant.
III. The Frontier
There is a boundary in every creative domain between what remains tacit — felt, intuitive, carried in the body or the gut or the accumulated instincts of a career — and what has become explicit enough for a machine to reliably act on. That boundary is where all AI-assisted creation lives.
Call it the spec frontier — specification being the engineer’s word for what every creative field has always needed: a description precise enough to act on. Unlike a score or a line of code, a prompt is noisy — the same words can produce different results each time. The tools for crossing the frontier are still crude, closer to Laban’s early sketches than to his finished notation. But the frontier itself is real.
On one side sits everything you know but cannot yet say: the sense that a scene needs more tension, the instinct that a melody should resolve differently, the feeling that a design is almost right but subtly wrong. On the other side sits everything you have managed to formalize: the prompt, the ruleset, the style guide, the storyboard, the set of examples that define what “good” looks like in your specific context. Between the two is a gradient: raw taste that cannot yet be spoken, selection that gradually hardens into criteria, and explicit descriptions reusable enough that someone else could act on them.
The spec frontier is not unique to any one domain. A writer producing serialized fiction with AI finds that the models can generate prose endlessly. What they cannot generate is the world — the internal rules, the character logic, the constraints that make a hundred chapters cohere instead of drift. The world-building is the spec. The prose is its output. A software team discovers the same thing: the AI can write any function they describe, but the quality of the description is now the bottleneck. The spec is the real product. Everything else is translation.
The crucial thing about the spec frontier is that the description does not have to come first. Often the opposite is true. You generate something and the first reaction is not a thought — it is closer to a flinch. Something is off. You cannot say what. You adjust a word, regenerate, and the new version is wrong in a different direction. But each failure reveals a constraint you did not know you had. After enough iterations you can finally say what you want — but only because you saw, many times over, what you did not.
The filmmaker who generates seven hundred clips and selects fifty-five is not failing to specify. She is specifying through selection — each rejection sharpening the implicit standard, each acceptance revealing a preference she might not have been able to articulate in advance. This is a looser form than a written brief — harder to reuse, harder to scale. But each round of selection drags intention from the tacit side of the frontier toward the explicit. Selection that accumulates into criteria is specification. Selection that doesn’t is the slot machine.
This is how creative work has always operated. What AI changes is not the process but its speed and scale — and with that speed comes a new demand. The faster the machine can produce, the faster you need to know what you think. The spec frontier does not demand a complete brief before you begin. But it does require you to converge on one, because without convergence the machine will keep generating, and you will keep pulling the lever, and the output will never become the work.
IV. The Honest Case Against
The spec frontier framework has a seam, and it deserves honest pressure.
The strongest objection is that making intention explicit may not merely capture it. It may constrain it in ways that damage the work. Some of the most important art in every medium has emerged from the collision between intention and accident — the unplanned brushstroke, the recording error that becomes a signature sound, the structural surprise in a prototype that suggests a form no one imagined. Jackson Pollock did not specify his drip paintings. Jazz musicians do not specify their solos. If AI routes all creation through a bottleneck of explicitness, it risks systematically eliminating the unspecifiable — the emergent, the accidental, the work that exists precisely because no one planned it.
Grimes’s frustration, looked at this way, may not be a complaint about tool quality at all. It may be a complaint about the kind of explicitness the tools demand — a legibility that kills the thing she valued most, the weird collision, the happy accident, the sound that emerged from not knowing exactly what she wanted. The spec frontier framework says the bottleneck is description. But what if some of the best creative work depends on the refusal to specify? What if the frontier, pushed too far outward, flattens the territory it claims to map?
There is a second pressure, political rather than aesthetic. The ability to specify — to articulate intention in precise, structured, abstract terms — is not evenly distributed. It correlates with formal education, verbal fluency, comfort with systematic thinking. If the human role in AI-assisted creation shifts from making things to describing things, the people who thrive will be those who were already fluent in abstraction: the educated, the verbal, the systematizers. The dancer who creates through the body, the craftsperson who works through the hands, the musician who plays by ear — these creators may find themselves on the wrong side of a new literacy divide. Laban’s notation, after all, did not serve every dancer equally. It served the ones who could read.
This is not a new pattern. Every previous technology that formalized creative intention created a similar divide. Musical notation split the composer from the performer and made the composer the higher-status role. Programming languages created a class of people who could instruct machines and a vastly larger class who could not. The spec frontier may be the latest version of this old rift.
But there is a counterweight worth noticing. AI also lowers the barrier to formalization. Natural language is easier than musical notation. A prompt is more accessible than a blueprint. A conversation with a model is less forbidding than a programming language. The frontier may be simultaneously a class divide and the lowest-barrier version of that divide in history — more people than ever can cross it, and crossing it matters more than ever. But the distance between writing a prompt and writing a description that actually works may reproduce the same advantages that formal training has always conferred. The barrier is lower. The skill ceiling has not moved. Whether that net effect is democratizing or concentrating depends on who builds the tools, how they price access, and what they optimize for — decisions that look technical but carry political weight.
The spec frontier is not “spec first.” It is spec somewhere.
Every previous technology that externalized creative intention revealed something hidden inside the act of making: that description was always the work. The making just obscured it. AI is doing this again, across every creative domain, at a speed that compresses what used to take centuries into years.
The spec frontier is moving outward. More of what once required tacit human presence becomes describable, executable at a distance, reproducible without the original creator in the room. But the frontier does not reach the horizon. There is always a residue of intention that arrives before the language for it does. New ideas begin as ambiguity. Taste resists formalization at its edges. The felt sense that something is wrong precedes, sometimes by a long time, the ability to say what.
Laban’s notation did not replace the dancer. It revealed what the dancer had always been doing — translating between something felt in the body and something a body could execute. The notation made that translation visible — first across the distance between rehearsal halls, now across the distance between what a creator means and what a machine can act on. It did not complete it. No notation ever does. The gap between intention and specification is not a problem to be solved. It is the permanent condition of creative work — the space where judgment lives, and the reason human presence in the act of making has survived every previous automation of execution.
AI is the most powerful execution technology ever built. What it leaves exposed — not as a consolation but as a structural fact — is the same gap Laban faced in that rehearsal hall: the distance between knowing what you want and being able to say it precisely enough that something else can bring it into the world.
Execution was never the whole craft. It was the part of the craft we could see.
That distance is the spec frontier. It is moving. It is not closing.