You Do It. You Decide.

Every semester I watch smart people use AI the way they were taught to use everything else: as a tool you operate. They write the prompt like a recipe. Open this, search for that, make a table with these five columns, use a professional tone, keep it under a page. Then they read what comes back, copy it somewhere, and do the next step themselves.

That works. It is also the slowest, smallest way to use the most capable thing you will ever have access to. So I teach two prompts, and I teach them before anything else.

• You do it.
• You decide.

Both prompts rest on one idea. Get out of its way. Stop telling it how to do something. Describe the state change you want, meaning what will be true when this is finished that is not true right now, and let it figure out the how.

School trained you to show your work. Every assignment rewarded the process: the steps, the outline, the method. So when you sit down with an AI, the instinct is to hand it a process. That instinct caps the result at your own skill. If you write the steps, the AI can only be as good as your steps, and you have made yourself the bottleneck in a system that is faster than you.

A state change is different. It describes the world after, not the path there. “I have three apartments I can afford, within walking distance of work, with viewing times booked for this week” is a state change. “Search Zillow and make a spreadsheet” is a step. The first leaves room for the AI to find a better path than the one you would have picked. The second forbids it.

Notice something about the right-hand column. It is not vaguer. It is more precise about the only thing that matters, which is what done looks like. It also gives you something the recipe never does: a test. When the work comes back, you check it against the state change. Do I have three apartments with viewings booked, or not?

This happens constantly. You describe an outcome and the AI returns a to-do list addressed to you. It is being polite. Most people ask AI for advice, so advice is its default. But advice is the work handed back to you with a bow on it.

When it tells you what you should do, answer with the first prompt. You do it. Not rudely, just plainly. Most of the time it simply does it, and you learn that the list was never a limit on what it could do. It was a guess about what you wanted.

The habit you are building is refusal. Every time you accept the list and start working through it yourself, you teach yourself that the AI is a consultant. Every time you push back, you learn what it actually is.

The second way it hands work back is subtler. It asks you questions. Would you like a formal or casual tone? Should I use option A or option B? Do you want this as a table or a list? Every one of those questions sounds respectful. Most of them are decisions you are not better placed to make than it is.

So the second prompt: You decide. If a competent person who understood your goal could make the call without you, let the AI make it. You can always redirect after you see real work, and redirecting a finished draft costs you far less than refereeing a hypothetical.

The US Army figured this out long before language models. Its doctrine on mission command defines the commander’s intent as a short statement of the purpose of an operation and the end state the commander wants, written so that subordinates can act without further orders even when the plan falls apart. The same doctrine describes mission orders as directives that stress the results to be attained, not how to attain them. That is exactly the stance I want you to take. You own the intent and the end state. The AI owns the path.

There are still decisions that belong to you. What you actually want. How much you are willing to spend. Anything sent under your name to another person. Anything you cannot undo. Keep those. Hand over everything else.

Sometimes you push back and it still cannot do the thing. Good. That is not a failure. That is the most important information you will get all week, because almost nobody knows where the edge is.

Researchers from Harvard Business School, Wharton, MIT and Boston Consulting Group gave that edge a name in 2023: the jagged technological frontier. In a pre-registered experiment with 758 BCG consultants, people using GPT-4 on 18 realistic tasks inside the frontier finished 12.2% more tasks, worked 25.1% faster, and produced work graded more than 40% higher in quality. On one task deliberately chosen to sit just outside the frontier, the consultants using AI were 19 percentage points less likely to get the right answer than consultants working alone.

The frontier is jagged. Two tasks that look equally hard to you can sit on opposite sides of it, and nothing on the screen tells you which side you are on.

That second finding is the warning inside this lesson. “It can’t” does not always look like “I can’t.” Sometimes it looks like a confident, well-written, wrong answer. This is why step one matters so much. If you described the state change, you have a test, and the test catches the confident miss. If you only described steps, you will check that the steps were followed, and they will have been.

When it genuinely stalls, sort the reason into one of three buckets, because only one of them is a real limit.

• No access

  It cannot reach your email, calendar or files. That is a setup problem. Connect the tool and ask again.

• Not permitted

  It will not do something on purpose, like spend your money without asking. That is a boundary, and often a good one.

• Not yet

  It has what it needs and still cannot produce the outcome. This is the frontier. Write it down.

Here is why the Not Yet bucket is so valuable: it expires. The frontier does not sit still, and it has not been moving slowly.

From four minutes to about seventeen hours in three years. METR’s own caveats are real: the tasks lean toward software and research, the intervals are wide, and critics argue the trend is sensitive to which tasks are in the suite. METR itself says it is. But the direction is not in dispute, and METR’s companion work across other domains, including math, computer use and robotics, found broadly similar rates of improvement.

So when you hit a wall, you are not learning “AI can’t do this.” You are learning “AI couldn’t do this on this date.” That is a far more useful sentence. Keep a simple list:

Date

The day you hit the wall.

State change I asked for

The outcome, written the way you wrote it.

What happened

It refused, it stalled, or it confidently got it wrong.

Retest

A date a month out. Ask again, same words, and note what changed.

Within a few months most students find items moving off that list. The people who notice first, who know this month what it could not do last month, are the ones who get to use it first. In a job market where everyone has the same tools, knowing where the edge is today is an advantage almost nobody is tracking.

I tell my students to let their first employer be their first customer. The way you become worth that is not by being the person who operates the tools best. It is by being the person who can describe outcomes precisely, hand them off without flinching, and knows exactly where the machine stops. Those are three management skills, and these two prompts train all three.

You do it, because the work is not yours to do anymore. You decide, because the call is not yours to make anymore. What stays yours is the intent, the end state, and the judgment to check that it happened.

Already true

In my classes, current assistants almost always complete work they first offered as a list of steps, if you simply ask them to. The capability frontier is uneven, and working outside it with AI can make results worse, not better. The length of tasks agents can finish on their own has been doubling in months, not years.

What has to happen

You have to break the recipe habit school gave you, write outcomes you can check, and actually push back instead of quietly doing the list yourself. And you have to keep the Not Yet list, because an edge you do not write down is an edge you will not notice moving.

Where I am probably wrong

“You do it” is the wrong answer for anything irreversible: sending, spending, signing. There you review first, every time. The Not Yet list also assumes the trend continues, and it could slow. If it does, some items will sit on your list for a year instead of a month. The lesson still holds, you just retest less often.

Four words. Say them more than feels polite. You do it. You decide.

This is how Charleston AI works with local businesses: we describe the outcome and deliver finished work, not tools and training. If you would rather receive the result than learn the steps, start at ai-chs.com.