The Prompt Is Not the Intent

Many projects do not begin as specifications. They begin as an internal sense: a feeling that something should exist, a frustration with what is currently available, a taste for a certain kind of interaction, a vague image of a tool or an essay or a system.

At the beginning, this sense may be genuinely hard to explain. That does not mean the person "doesn't know what they want." They know something. They just do not yet know it in language.

They know it through reaction. They see one version and feel no, not that. They see another and feel closer. They notice that a technically correct solution feels too heavy, or that the important part was never the feature they asked for but the experience around it.

This reaction is not secondary feedback. It is how the intent becomes discoverable at all. The interior becomes clear through contact with exterior forms.

Ken Wilber's four-quadrant model gives this a useful shape. You do not need to buy all of Integral Theory for the distinction to help. The model splits experience along two axes: individual and collective, interior and exterior.

• Individual interior: intent, taste, felt sense, what the creator wants, values, fears, senses.
• Individual exterior: visible behavior, prompts, clicks, edits, choices.
• Collective interior: shared meaning, brand, trust, team assumptions, culture.
• Collective exterior: systems and structures, architecture, workflows, databases, organizations.

AI tools operate through exterior artifacts, and they are unusually fluent in the collective exterior. A model has absorbed the aggregate of what everyone else has already built: the visual conventions of websites, the structural patterns of applications, what "clean" and "modern" and "professional" have tended to look like across thousands of examples. This is a real strength. It is the collective pattern language an individual designer would need years to internalize.

But the strength has a shadow side. When your individual interior has not yet resolved, the AI does not sit with the ambiguity. It defaults to the dimension it knows best.

Ask for "a beautiful website" without yet knowing what beautiful means for this project, and the system reaches for the statistically safest rendering of what beautiful websites have tended to be. Sometimes it will not even pick: it ships a light theme, a dark theme, and a toggle between them so that no commitment has to be made.

It has not guessed wrong. It has avoided guessing, and built that avoidance into the product as if it were a feature.

To be fair, the quadrants are entangled by nature. Aesthetic judgment is never pure individual taste, sealed off from the world. What you find "clean" or "premium" was already shaped by the collective visual vocabulary of every site you have used.

The problem is not that AI draws on the collective. The problem is that when the individual signal is unresolved, the collective average quietly stands in for it, and nobody notices that a substitution has taken place.

This is not only a philosophical point. It also shows up in software requirements research. In a 2022 controlled study, Ferrari, Spoletini, and Debnath had analysts run interview-based elicitation sessions and then traced the resulting requirements back to the customer's initial ideas. Only 30% to 38% of the requirements could be fully traced to those initial ideas. More requirements emerged later, when analysts browsed similar products in app stores. The authors' conclusion is almost the thesis of this essay in miniature: requirements are not elicited in the strict sense. They are co-created, with the analyst playing a crucial role.

The initial idea is rarely the whole requirement. Something happens through conversation, comparison, prototyping, and reaction: the person sees possibilities, the collaborator offers interpretations, the idea evolves.

Yet most AI workflows still behave as if the requirement were fully contained in the first prompt. If requirements are co-created, then the early phase of an AI interaction should be optimized for discovery, not generation. It should stay open long enough for the person's own signal to surface, rather than being closed by whichever party moves first to commit.

Brene Brown has a delegation tool in Dare to Lead called "paint done": instead of handing someone a vague instruction, you describe what done looks like in enough detail that hidden expectations and unsaid intentions become visible.

• "Make it better."
• "Keep it simple."
• "Make it feel premium."

Every one of those hides a dozen assumptions the receiver is forced to guess at. When the receiver is an AI and the individual signal is unresolved, the guess often defaults to the collective average.

The detail of Brown's tool that matters most here is the phrasing: it is "let's paint done." The picture is painted together, by the asker and the receiver. Which means it is the same finding as the requirements study, arrived at from leadership practice instead of software research: done-ness is co-created, not extracted.

Creative work adds a twist, though. Often the person giving the instruction cannot paint done yet, even collaboratively. They need to see several possibilities before they can say what done looks like. This is exactly where AI could become far more useful than it currently is, not rushing to produce the final artifact, but helping the person discover what "done" means.

A prototype is not just an early version of a product. It is a mirror for intent. Seeing one teaches you what you actually want: the main action is wrong, the page has too many choices, the feature you thought was central turns out to be secondary. Something external finally gives your internal sense something to push against.

In this sense, feedback is not correction. It is revelation.

"This feels wrong" sounds vague, but it is data. Interior data. A good creative system would investigate it rather than dismiss it.

• Wrong how?
• Too heavy?
• Too generic?
• Too much like everything else?
• Not enough like the original feeling?

The point is not to turn every feeling into a metric. The point is to treat the feeling as part of the system.

This is where the shape of AI-assisted creation needs to change, not just the philosophy behind it. When a system commits to every layer of a project at once, such as structure, visual design, tone, copy, and edge cases, in a single pass, it forecloses your ability to react. There is nothing left that is cheap enough to safely disagree with.

A fully polished version that turns out to be wrong is expensive to unwind. The tokens spent on styling and detail were spent before anyone knew whether the direction was right. Worse, its finish suppresses objection. Pushing back on something that presents itself as done carries more friction than pushing back on something visibly provisional, so people accept what they should have challenged.

A bare, structural version is different: the core flow, the basic shape, no polish. It is cheap to build and cheap to discard. Four bare variations will often cost less than one fully realized version that turns out to be wrong. Unlike the wrong finished version, each bare one earns its keep as a mirror.

Bare does not have to mean unstyled, either. It means committing to one layer at a time. If the open question is the flow, show skeletons. If the open question is the tone, show cheap tone probes: a styled fragment, a mood, not a finished product. Keep whichever layer is unsettled cheap, and let the other layers wait their turn.

Iteration is not expensive by nature. It becomes expensive when each iteration insists on being complete.

The next generation of AI tools should not just generate faster. They should help triangulate intent by treating the first prompt as a starting point rather than a source of truth, and by resisting premature completeness, because completeness is what closes off reaction.

A better workflow would open a possibility space and keep it deliberately bare: three structural interpretations of the idea, a version that exaggerates one direction, a version that shows what the project should not become.

Then you react:

• "That one is closest."
• "The structure is right, but not the tone."
• "I did not realize this was actually about trust."
• "This should feel more like a tool than a product."

The system should treat those reactions as first-class information, not as annoying corrections. They are how the hidden intent becomes visible.

Concretely, this means AI-assisted creation needs an explicit intent layer: a living brief that sits above the requirements and the code, which the AI drafts, maintains, and revises with every reaction. Not a static document, but a running model of what this thing is trying to become, why it matters, what it should feel like, and what would make it wrong.

In practice it might sound like this:

The loop is simple. Interior sense becomes prompt. Prompt becomes bare prototype. Prototype provokes reaction. Reaction refines intent. Refined intent shapes the next artifact. Only once the intent has settled does the artifact earn the right to become expensive.

And you can run this loop today. No tool enforces it yet, but nothing stops you from doing it manually. Before letting an AI build anything complete, ask for three bare variants: structure only, and tell it explicitly not to style or polish. React in "closer" and "further" language rather than instructions. Then ask the AI to restate what it now believes you want, in its own words, before it builds the next round. You are maintaining the intent layer yourself, in conversation. It costs minutes, and it saves the long fight with a wrong artifact that got expensive too early.

Without an intent layer, AI systems move too quickly from vague desire to concrete artifact. The user is still discovering what they mean while the system is already implementing what it guessed, every layer at once. The result is not just bad output. It is premature concreteness: the artifact becomes heavy before the idea becomes clear.

The frustration it produces is not "the AI did not follow instructions." It is: "the AI externalized my words, but not my meaning."

The prompt is an artifact. The prototype is an artifact. The requirement and the code are artifacts. They are external expressions that help reveal, test, and refine something internal, but they are not the internal thing itself.

If AI tools are going to become real creative partners, they need to stop treating the first external expression as the whole truth. And they need to stop reaching for their greatest strength, fluency in what everyone else has already built, as a substitute for the individual answer they have not yet found.

The future of AI-assisted creation is not only faster generation. It is a better bridge between internal experience and external form.