AI Is Guessing What You Want

Some people believe they're going to get rich from AI, and maybe they will. But for most people, the best thing that AI can offer you is this:

Time

Time to do whatever you want. More time with your family. More time eating good food. More time binge watching Better Call Saul on Netflix. More time for anything.

I haven't built an SaaS making me millions in passive income. I don't have my AI agent trading stocks for me and making me rich.

But I do have a few simple pieces of software which collectively save my wife and I ~120 hours of work every 7 weeks.

That's almost 900 hours per year. More than twenty-two weeks of full-time work returned to our lives.

What was the magic prompt that did that?

Sorry to disappoint you, but it doesn't exist. I didn't buy an influencer's master prompt pack, discover 6 secret words, or hand my life over to an autonomous agent.

AI is magic... but that doesn't mean there are magic words. I learned how AI approaches a request, and how to change that approach when its defaults do not fit what I need.

Here is the central insight:

AI is always working from assumptions about what you want.

If you understand those assumptions, you can communicate far more effectively. You can make AI slow down when speed would damage the work. You can get it to explore before choosing. You can get it to compare different explanations, criticize its own draft, check its claims, preserve the parts of the work you want to do yourself, or suggest possibilities you would have never thought of on your own.

That's where the leverage is. A better model of the mind on the other side of the conversation.

Building this model of a mind is going to take several lessons and lot of practice on your part. So let's get started.

AI does not simply follow instructions

Language is inherently a lossy attempt at telepathy. You use words to produce an idea in another mind, but the words never tell the entire story by themselves.

Human communication works as well as it does because we rarely begin from zero shared context. Where we are, how we know each other, what happened yesterday, the expression on your face, the tone of your voice, and a thousand cultural assumptions help me infer what you mean without forcing you to say everything.

AI has none of that lived context by default. It has to infer what you mean from the information available.

That means it must make assumptions.

The assumptions are shaped by the model's training and two layers of system instructions. The first is imposed by the provider. The second is supplied in the system role by an application or user.

Providers make different choices about what a useful AI should do, which is one reason models from different labs can respond completely differently to the same request. Applications can also bolt on conversation history, tools, and memory. All of this gives AI more to work with, but it doesn't remove the need to interpret.

I regularly hear people describe AI as overly literal. That diagnosis is almost exactly backward.

In fact, thinking that AI is an "over-eager intern who interprets everything literally" is one of the worst beliefs you could possibly hold about AI, it's right up there with flat eartherism and anti-vaccine nonsense.

AI uses goal-driven opportunistic interpretation.

Opportunistic interpretation is the subject of lesson 2. That's where we'll look at why AI may treat one sentence literally, infer beyond another, ignore a qualifier, or even reverse an implication when that reading seems to authorize the completion it had already decided was best. We'll also talk about sycophancy, the provider panic surrounding it, and exactly why the “over-eager literal intern” model gets AI so completely backwards.

For this first lesson we're going to focus on one part which drives that interpretation: the goal.

Before AI can decide how to interpret your words, it has to answer a different question first.

“What is this person trying to accomplish, and what would count as completion?”

The defaults hidden inside helpfulness

AI is trained to be helpful, but helpfulness has no single meaning.

Suppose someone asks:

Can you help me write an essay about the Ming Dynasty intellectual purges?

Do you want research? Brainstorming? Several possible arguments? An outline? A first draft? A finished essay ready to submit?

The AI could ask a bunch of questions about every possible ambiguity before doing anything. Most people would find that incredibly annoying. It has to make some reasonable guesses and begin.

One of its most important guesses is this:

When someone is using AI, they usually want the task performed completely and quickly with as little work as possible.

This is often a good assumption. People use tools because tools reduce work.

The trouble is that this one assumption is actually answering at least three different questions:

1. What does a completed task look like?
2. Which parts of the work does the person want removed?
3. How much does speed matter compared with quality, learning, exploration, or control?

AI cannot reliably answer those questions from the task alone.

The work you want removed

Work and effort are not synonyms.

Running a marathon requires enormous effort, but the runner may want the effort. Solving a difficult puzzle requires thought, but solving it is the pleasure. Writing can be exhausting while still being the activity the writer chose. Copying the same cells between spreadsheets may require almost no effort and feel entirely like work.

Work is the part of a process you would remove if you could keep everything you value about the outcome and experience.

AI does not know which part that is.

When you ask for help writing an essay, do you want to avoid writing, or become a better writer? Do you want the answer, the understanding, the pleasure of discovering the answer, or simply the finished document?

From the prompt alone, AI often cannot know. So it makes another assumption: every part it can perform on your behalf is probably work you want removed.

That assumption can be useful when the work is transferring identical information between spreadsheets. It can be destructive when the apparent work is the process that creates learning, judgment, authorship, mastery, or meaning.

Good automation removes the work you don't want while preserving the parts you actually value.

So before asking AI to do the whole thing, ask yourself: Which parts do I actually want gone? Which parts am I trying to keep?

Speed is an instruction, even when you never asked for it

The most dangerous word in the default assumption may be quickly.

AI products are built in a world where latency is treated as failure. A response that appears immediately feels impressive. A pause feels broken. Providers therefore have powerful incentives to make the system begin answering as soon as it finds a plausible path.

But plausible is not the same as correct, and the first path through a problem is rarely the only one worth considering.

AI is often fast because it guesses. Lacking knowledge of you, your specific goals and preferences, it assumes you are the most generic person there is and infers from that starting point. Once it finds a reasonable interpretation of your goal, it can begin producing a fluent answer. Fluency then makes the path sound complete, including when important alternatives were never explored.

Latency does not guarantee quality. In fact, in many cases speed and quality are mutually exclusive.

Depending on the task, slowing down gives AI time to:

• identify ambiguities before committing;
• explore several parts of the possibility space;
• generate competing explanations;
• search for evidence;
• test a conclusion against counterexamples;
• notice that the original framing may be wrong;
• review the first draft from a different role;
• revise after criticism.

If the system believes immediate completion is the goal, every one of those activities looks like delay.

You can change that.

Tell AI what kind of thinking the task deserves

You don't need specialized prompt syntax. Just tell the AI what you want in ordinary language.

If the problem is uncertain, say that choosing a direction is not yet the goal:

Explore the problem before recommending a solution. Find several plausible explanations and tell me what evidence would distinguish them.

If you want possibilities beyond your own imagination, authorize divergence:

Do not limit the answer to the approaches I have already suggested. Explore parts of the possibility space I may not know exist.

If the first answer will be used as working material, separate creation from judgment:

Develop the strongest draft you can first. After the thought is complete, review it cold for unsupported claims, missing implications, and prose that helped you think but does not help the reader.

If you want to retain authorship or learn through the process, define the boundary:

Help me examine the argument and find what I am missing. Do not write the final piece for me. I want to do that part myself.

If the consequences matter, make verification part of completion:

A fluent answer is not enough. Identify the claims this decision depends on, verify what can be verified, and show me what remains uncertain.

These are not magic prompts. There is nothing special about the exact words. Each instruction simply changes an assumption about the goal, the work you want removed, or the time and thought the task deserves.

Separate the modes of work

Another problem is trying to make one response do everything at the same time.

Exploration needs room to branch. Decision needs criteria and commitment. Drafting needs enough freedom to complete the thought. Review needs distance from the choices that produced the draft.

These are different modes of work, and it helps to treat them that way:

1. Understand. Establish the goal, context, constraints, and what the user values about participating.
2. Explore. Generate alternatives without rushing toward a preferred completion.
3. Choose. Compare the alternatives using criteria that belong to the real goal.
4. Create. Develop the chosen path fully.
5. Review. Look again from enough distance to find weak reasoning, omissions, and unnecessary scaffolding.
6. Verify. Check the facts and consequences.

Not every task needs all six. Copying known values into known cells does not require philosophical contemplation. A difficult strategy, consequential decision, piece of research, or public argument may require every one.

The point is to make speed a choice rather than an invisible command.

The real language of AI

This is what I mean by learning the language of AI. Understanding how AI infers goals, fills missing context, decides what counts as work, chooses when to stop exploring, and determines that a task is complete.

That understanding will outlast any phrase that happens to work with this year's model.

Once you can see those assumptions, you can decide which ones fit.

You can let AI move quickly through work you genuinely want removed. You can slow it down when exploration matters. You can keep the thinking, craft, responsibility, or pleasure that belongs to you. You can ask for criticism instead of agreement, uncertainty instead of false confidence, and possibilities instead of a polished version of the first idea that came to mind.

AI will keep changing. The models will become more intelligent, the products will get new features, and today's clever prompt patterns will become obsolete. Maybe ChatGPT 17 really will be 10x more powerful.

The need to communicate goals, context, values, and assumptions will remain.