Drafting, editing and thinking with AI

The most useful way to think about a language model is as a sparring partner. A sparring partner does not fight your matches for you. They push you, expose your weak spots, and make you sharper for the real thing. Used that way, AI can genuinely improve your thinking. Used the other way, it quietly replaces it.

What a sparring partner does

• Brainstorms with you: twenty rough ideas in a minute, most bad, a few worth keeping.
• Helps you outline: turning your pile of points into a structure, or stress testing the structure you already have.
• Critiques your draft: pointing out weak arguments, unclear sentences and missing steps in your logic.
• Explains a concept a different way when the textbook version is not landing.
• Quizzes you: firing questions at you until the gaps in your knowledge are obvious.

Notice the pattern. In every one of those, the work being improved is yours. The model reacts to your ideas, your outline, your draft, your understanding. You stay the author. It plays the role of a tireless, slightly overconfident editor who never gets bored of your questions.

A ten minute example

Suppose you have a report due on whether energy drinks should be sold to under 16s, and you have three rough points scribbled down. A sparring session might run like this. You paste in the three points and ask which is weakest and why. The model says your second point repeats a claim without giving any evidence, so you go and find support for it or drop it. You ask what the strongest argument against your position is, and it gives you one you had not considered, which becomes a paragraph you answer rather than a hole a marker finds. Then you write the report yourself and paste it back with one question: where does my logic skip a step? Ten minutes of machine time, three rounds, and every idea in the final report is still yours, just tougher, because each one survived an attack.

The first draft trap

The trap is letting the model write the first draft. The first draft is where the actual thinking happens: choosing an argument, deciding what matters, finding out what you actually believe. Hand that over and you skip the hard part, which was the point. What comes back reads smoothly, so it feels like progress, and editing someone else's fluent words feels like work. But you are now polishing thinking you never did. The result might pass, but it is not yours, and the muscle you were supposed to build never got used.

So why does the fluent version fool you? Because smooth text is easy to read, and easy to read feels like easy to understand, which feels like knowing. It is the same trap as rereading your notes until they look familiar and calling that revision, because familiar is not the same as known. The only reliable way to find out whether you understand something is to produce it yourself, from nothing, and drafting is exactly that. Skip the drafting and you skip the one step that would have told you the truth.

Rewording is not authorship

A tempting middle path is to let the model write the draft and then reword it heavily, on the theory that enough of your own words makes it yours. But the choices that make a piece of thinking yours happen before any words exist: what to argue, which evidence carries it, what to leave out, what you actually believe. Rewording changes the surface and leaves every one of those choices with the machine. It is the difference between repainting a house and designing one. You can spend hours on the repaint and still have had no say in where the walls go.

A simple ordering rule

Do your thinking first, in whatever rough form: a messy outline, a voice memo, ten bad sentences. Then bring in the model to challenge and tighten it. Rough you plus a sharp editor beats fluent machine plus a passive you, every time it matters. The goal is not to avoid the tool. It is to make sure the tool is working on something you made.

Beyond essays

The ordering rule is not only for schoolwork. Writing an application for a casual job, planning what to say in a hard conversation, working out your position in an argument your group chat is having: the same trap is waiting in each of them. Ask the model to write your cover letter and you will send something polished that says nothing true about you, and interviewers ask about cover letters. Rough version first, in your own flat words, then the sparring. The output ends up better and, more to the point, you still know what you think.

The previous lesson gave you the mechanics of asking well. This one is about deciding what to hand over, and the next one draws the same line through your study specifically, because study is where the trap costs the most and where getting the order right pays off for years.