Some Prompts Should Deliberately Leave Room

Sometimes I give an Agent an article, a conversation, or even a pile of unorganized material and ask just one question:

What is this really about?

By the standards of prompt engineering, this question fails almost every test. It assigns no role, prescribes no steps, defines no output format, and does not even clarify what “this” refers to.

The Agent has to struggle with it too. It does not know whether I want a summary, a critique, or a distilled argument. All it can do is try to work that out for itself: What does the author actually care about? Which terms are merely packaging? Is there a thread connecting the different pieces of material? If everything expendable were removed, what judgment would remain?

Strangely, prompts like this sometimes produce more interesting answers.

Not because ambiguity is inherently superior to clarity, but because some tasks are difficult precisely because we do not yet know how to state the question clearly.

Suppose I begin with this instruction instead:

Summarize this article in five sections: background, problem, causes, solution, and lessons.

The Agent will certainly produce a more consistent result. But I have also decided in advance what the article contains. Its job is now simply to fit the material into five boxes. Even if the author is not telling a “problem–cause–solution” story at all, the Agent can still organize the text as though it were one.

The more specific the prompt, the greater the control. But that control may also conceal the questioner’s blind spots.

We often think of a good prompt as a complete blueprint: the objective, process, format, and standards are all specified in advance, and the Agent only needs to follow the plan. This works well when we already know roughly what the answer should look like—changing a piece of code, extracting fields from a contract, or arranging data in a table.

But some tasks call not for construction, but for exploration.

Reading an unfamiliar article, deciding what problem a product truly solves, understanding why a debate keeps missing the point, or even figuring out what vaguely bothers us—these tasks do not yet have maps. When a prompt is too complete, it often means that before the exploration has begun, we have quietly drawn a false map.

“What is this really about?” sounds vague, but it hands an important part of the work to the Agent: not only answering the question, but also trying to define it.

What matters here is not ambiguity itself, but the reversibility of exploration.

The Agent can return with an initial interpretation. If I am not satisfied, I can explain what is wrong: it follows the author too readily; it restates the argument without explaining why it matters; it identifies the surface claim but misses the interests underneath. The Agent can then try another direction. Only after several rounds may we gradually realize what we were trying to ask in the first place.

A prompt is no longer a specification that must be correct on the first attempt. It is more like an initial probe: push it in, see what is there, and adjust the direction in response.

But “just ask it to revise the answer” is not as easy as it sounds.

The first answer creates an anchor. Once the Agent interprets the material as a problem of efficiency, even a request to go deeper may only lead it to search for deeper causes of inefficiency. People do the same. Once a sufficiently fluent explanation appears, we easily forget that it is only one among many possible explanations.

So when the first answer takes the wrong direction, saying “go deeper” usually does little good. It is more effective to make the Agent genuinely leave the original frame:

Set aside the previous explanation. Offer three competing interpretations. Explain what each one accounts for and what each one leaves out.

Revision continues along the same road. Reinterpretation chooses another road at the fork.

Of course, not every task is suited to this approach.

If an Agent is changing a production environment, sending a customer email, giving medical advice, or executing an irreversible transaction, we cannot let it explore freely and fix the result afterward. The first result may already carry a cost. The objective, permissions, boundaries, and acceptance criteria must be explicit in advance.

Deliberate ambiguity is appropriate for tasks that are cheap, reversible, inspectable, and within our ability to judge. Writing, reading, research, and product exploration often belong in this category. Execution, commitments, and high-risk decisions usually do not.

I now prefer to distinguish between two things: leaving the path unspecified and leaving responsibility undefined.

A good open-ended prompt can avoid telling the Agent how to think while still giving it enough material. It can allow surprising answers while requiring evidence. It can postpone defining the correct answer while making competing interpretations test one another.

For example, we might add one sentence after “What is this really about?”:

Do not rush to summarize the author on their own terms. First propose several possible core claims and test each one against the material. Then tell me where you are least certain.

The prompt still leaves room. But that space no longer comes from laziness. It has been reserved for exploration.

We have long worried that Agents will misunderstand our intentions, so we keep making prompts longer. But sometimes the greater risk is that an Agent will execute with perfect accuracy an intention that we articulated too early and had not yet understood ourselves.

A prompt does not always need to contain the shape of the answer in advance.

Sometimes it only needs to place the Agent in front of the problem and say:

Go in and take a look.