Intent Before Execution

The principles behind Paper

Artificial intelligence is becoming increasingly capable of execution. AI systems can write code, analyze documents, search large bodies of information, generate designs, operate tools, and carry out increasingly complex sequences of work with less direct human involvement.

We think this progress is useful. We also think it creates a problem that is not primarily about the capability of the AI itself: what happens when the person using the system does not fully know what they want yet?

Meaningful work rarely begins with a complete specification. It often begins with fragments: an idea, a constraint, an observation, something that previously failed, several competing requirements, or simply a sense that something is wrong without a clear understanding of why. People frequently discover what they mean through the process of working through a problem.

Increasingly, however, we are giving that unfinished thinking directly to systems designed to produce answers or take actions.

Paper is built around a working theory: there should be a layer between unfinished human thought and AI execution.

We call that layer intent formation.

This is a theory, not established research

The ideas described here have not been scientifically validated. We have not conducted controlled studies demonstrating that intent formation improves AI outcomes, nor do we have a proprietary dataset proving that existing AI workflows systematically fail in the ways described below.

This is not a research paper, and these ideas should not be presented as established findings.

They are working hypotheses based on observations made while building and working with AI systems, mistakes encountered during that process, and questions that we believe are worth exploring. Some of our assumptions may be wrong. The need for a distinct intent-formation layer may diminish as models improve. Other approaches may solve the same problem more effectively, or the problem may prove important only for certain kinds of work.

Paper is, in part, an attempt to find out.

The problem we see

Consider a simple instruction:

The settings page feels messy. Clean it up, but keep the important things easy to reach.

There is not enough information in that statement to determine a single correct course of action. What does “messy” mean? Which settings are important? Does “clean it up” mean changing the visual hierarchy, reorganizing navigation, removing options, or hiding advanced settings?

To proceed, an AI system has to interpret what the person meant.

Modern models can be remarkably good at making these interpretations. That capability is part of what makes them useful. The concern is not simply that an AI might make a bad inference. It is that a plausible inference can be accepted so naturally that neither the person nor the system clearly distinguishes it from an explicit decision.

An incomplete instruction can lead to an AI inference. That inference can become an assumption within the work. Later decisions can then build on that assumption until something the person never explicitly decided begins to function as an established requirement.

The inference itself may have been entirely reasonable. The distinction we care about is simpler: a reasonable inference is not necessarily the same thing as human-established intent.

Why we do not think better prompting completely solves this

Clear instructions generally produce better results than unclear instructions, and prompt engineering can be useful. However, the advice to “write a better prompt” contains an assumption of its own: that the person already understands their intent well enough to specify it.

Sometimes they do. Sometimes determining what the instruction should be is the problem itself.

A person may know that a system feels wrong without knowing what architecture should replace it. They may have several competing product requirements without having decided which one matters most. They may understand the desired outcome while remaining uncertain about important constraints. They may even change their mind while explaining the problem.

In those situations, producing a better specification is not necessarily the first task. The first task is developing the thinking from which that specification can emerge.

Intent formation happens before specification.

Why we do not think more context completely solves this

AI systems are increasingly being given larger context windows, persistent memory, project files, repositories, previous conversations, tool access, and external information. These capabilities can substantially improve what an AI system is able to understand and accomplish.

But having more context is not the same as having established intent.

A context window can contain explicit decisions, possibilities that were merely considered, obsolete decisions, AI-generated suggestions, unresolved questions, failed approaches, temporary assumptions, observations, and contradictions. All of these may be useful, but they do not necessarily have the same authority.

For that reason, we think there are at least two different questions involved. The first is whether the AI has enough information. The second is whether the system can distinguish information that was merely present or proposed from information the person actually established.

More context solves the first problem more readily than the second. In some cases, increasing the amount of context may make that distinction more important rather than less.

Why we do not think agents completely solve this

AI agents are becoming capable of carrying out increasingly complicated sequences of work. An agent can inspect a repository, develop a plan, modify files, run tests, diagnose failures, revise its approach, and continue working.

These systems address an important execution problem: given an objective, how can the system accomplish it?

Paper is concerned with a question that can arise before that: what should the objective actually be?

When execution is slow or expensive, there is a natural incentive to spend time deciding what should happen before acting. As AI makes execution faster and cheaper, it becomes increasingly easy to collapse thinking and execution into a single process: a person expresses an incomplete idea, and the system immediately begins acting on its interpretation.

Our hypothesis is that some forms of work benefit from preserving a deliberate stage between those two events.

Our theory: intent formation

We believe AI can participate in forming intent without being given sole authority to establish that intent.

An AI system can surface assumptions, contradictions, unanswered questions, alternative interpretations, missing constraints, possible consequences, uncertainties, and details that the person may have overlooked. These contributions can be valuable precisely because the AI can examine the problem from perspectives the person may not have considered.

But those contributions should remain proposals until the person decides what should become established context.

This creates a distinction between what the AI believes the person might mean and what the person has actually decided they mean.

Our shorthand for this principle is:

AI proposes. Human establishes.

That distinction is the central idea behind Paper.

The cost of additional friction

A reasonable objection is that AI is supposed to make work faster. Introducing an additional stage before execution appears to work against that goal.

In many situations, we agree.

A request to summarize an email, convert a measurement, explain a function, or answer a straightforward factual question probably does not need a dedicated intent-formation process. Not every interaction with AI requires deliberation, and we do not believe Paper's approach should be inserted into every AI workflow.

We are more interested in work where ambiguity can compound across subsequent decisions: designing systems, making product decisions, defining requirements, planning substantial changes, working through uncertain technical problems, maintaining long-running AI context, or delegating increasingly consequential work to execution systems.

Our hypothesis is that, in those situations, a small amount of deliberate friction before execution may prevent substantially more friction later.

Again, this is a hypothesis. We have not proven it.

“Modern models already understand what I mean”

Often they do.

Our argument is not that AI inference is inherently undesirable. Without inference, modern AI systems would be considerably less useful. The issue we are exploring is the boundary between inference and authority.

An AI system may correctly infer a person's intent repeatedly. The interesting case is what happens when an incorrect inference is plausible enough to become part of later work without being noticed.

As models become better at interpreting people, their incorrect interpretations may also become more convincing. A misunderstanding expressed fluently and plausibly can be harder to recognize than an obviously poor response.

Paper is therefore not attempting to eliminate inference. The goal is to make consequential inference visible enough that a person can accept it, reject it, refine it, or leave it unresolved.

“Why not just do this in ChatGPT, Claude, or another AI system?”

You can.

A person can instruct an AI system not to solve a problem immediately, ask it to challenge assumptions, request clarifying questions, have it describe what it believes the person means, and manually maintain a record of decisions. This can work very well.

Paper's theory is not that these behaviors are impossible in a conversational interface. The question is whether they become more reliable when the workflow itself represents the distinction between AI-generated interpretation and human-established context.

Instead of depending entirely on prompting discipline, that distinction becomes part of the environment in which the work occurs.

We do not yet know how valuable that distinction will prove to be. Testing that question is part of the reason Paper exists.

“Doesn't the AI still influence the person?”

Yes.

Paper does not make AI neutral. The questions a model asks affect what someone considers. The alternatives it presents can frame a decision. The assumptions it chooses to surface reflect the model's interpretation of the work. Even deciding what deserves attention can influence the person using the system.

Explicitly establishing a decision does not remove those effects.

Our claim is narrower: AI interpretation should not silently become human intent.

The person may still be influenced by the system, but the workflow attempts to preserve a meaningful boundary around who ultimately establishes the context that subsequent work should rely upon.

Intent does not stop at execution

Our initial thinking about Paper focused primarily on information flowing toward an AI system. Over time, we have become increasingly interested in the information that comes back from execution.

Execution produces knowledge.

Code fails. APIs behave differently than expected. Users misunderstand interfaces. Infrastructure exposes constraints. Previously reasonable assumptions encounter reality and turn out to be wrong. These outcomes can become valuable context for future work.

We do not think everything produced during execution should automatically become permanent context, either. The same principle applies in the other direction: the person should determine which outcomes matter enough to influence future decisions.

This suggests a broader loop:

Unfinished thinking → intent formation → human-established context → AI execution → real-world outcomes → human judgment → stronger context

The process can then repeat.

AI can participate throughout this loop. The important distinction is that the human determines what becomes authoritative.

What we are building

Paper is our attempt to build this missing layer.

It begins as a notebook because unfinished thinking needs somewhere to exist before it becomes an instruction. A person can write normally without requiring every thought to immediately become a prompt or an action.

When the person explicitly requests Review, AI can examine the work and surface questions, assumptions, contradictions, or other issues worth considering. Those suggestions can then be examined, rejected, developed, resolved, or incorporated into the person's established thinking.

The objective is not for Paper to produce the final answer on the person's behalf. The objective is to help transform unfinished thinking into better-established intent that can eventually be handed to systems designed to execute it.

In that sense, Paper begins upstream of AI execution. As we experiment with bringing meaningful outcomes from execution back into established context, the same model may also become useful downstream.

In either direction, the principle remains the same: the loop should remain human-directed.

What would prove us wrong

We believe this question is worth investigating, but we do not want to confuse conviction with evidence.

There are several ways our theory could turn out to be wrong.

AI models may become sufficiently capable of preserving ambiguity, distinguishing assumptions from decisions, and requesting clarification that a separate intent-formation layer provides little additional value. People may find explicit establishment more annoying than useful. The distinction between proposed and established context may matter only for a narrow category of work. Conversational interfaces or execution agents may develop mechanisms that address the same problem more naturally.

It is also possible that people simply prefer correcting AI after execution rather than spending additional time clarifying intent beforehand.

Those are not objections we need to dismiss. They are possibilities the product has to confront.

If experience shows that our assumptions are wrong, then the theory should change with the evidence.

Why build it anyway?

Because the direction of AI makes the question increasingly interesting to us.

Execution is becoming cheaper. Models are becoming more capable. Agents are becoming more autonomous. The amount of work that can occur between a human instruction and the resulting action is increasing.

Our hypothesis is that these developments do not make human intent less important. They may make it more important to distinguish what a system inferred from what a person actually established.

We do not yet know whether that hypothesis is correct.

Paper is how we intend to find out.

Intent before execution.
AI proposes. Human establishes.