There's a particular kind of disappointment that comes from reading a polished paragraph you didn't really think your way into. It might be correct, even publishable. But it leaves behind a faint sense of distance — like the work happened slightly outside of you. I know that feeling because I've invited it myself. Early on, I used AI mostly to produce language quickly. The results looked good enough and felt wrong enough that I had to admit something important: a strong sentence isn't automatically evidence of strong thought.

The way I use AI now is narrower and, honestly, more satisfying. I don't ask it to replace the work of forming a view. I ask it to help me pressure-test, organise, or widen my thinking after I've already started somewhere. That shift changed everything. Instead of outsourcing authorship, I started using the tool the way I'd use a good notebook or a careful editor — as something that helps me hear my own thinking more clearly.

Begin With an Unfinished Human Idea

The best AI conversations I have rarely start with a finished question. They start with something partial — a vague tension, a paragraph that feels almost right, a pile of notes that clearly belong together but haven't found their structure yet. I bring the tool an unfinished human idea on purpose. That preserves the asymmetry I want: I'm still responsible for the direction and meaning of the work. The model is helping me surface implications, not decide what matters.

This also protects me from one of the easiest traps in AI-assisted writing — asking the model what I think before I've done enough thinking to actually know. If I start too early, the output becomes strangely persuasive, especially when I'm tired. Beginning with my own partial draft gives me something to push against and something to return to. It keeps the whole conversation anchored in actual intention rather than generic fluency.

Ask Questions That Improve Judgment

The prompts I rely on most are the ones that sharpen judgment rather than generate prose. I ask for missing assumptions, possible objections, stronger contrasts, concrete examples, overlooked implications. Sometimes I ask the model to summarise my argument in one sentence so I can see whether I've actually made one. Sometimes I ask which part of a paragraph feels vague. These prompts are useful because they turn language into feedback rather than an end product.

There's a practical reason this compounds over time. Judgment is what sticks. If a tool helps you notice better questions, structure ideas more carefully, or test your reasoning more rigorously, that benefit stays with you after the session is over. Borrowed phrasing doesn't. It disappears the moment you're facing the next blank page.

Save the Tool for the Middle, Not the End

AI tends to be most useful in the middle of a project. Too early, and it can overwhelm fragile ideas with polished language before you've figured out what you actually want to say. Too late, and it might tempt you to polish away roughness that was actually doing something. In the middle, though, it can be excellent — that's when you know enough to direct the conversation but still need help seeing shape, sequence, and alternatives. I think of it as assisted deliberation rather than assisted writing.

This middle-stage use is especially valuable for longer pieces. After drafting a few sections, I'll often paste them in and ask: what's repeating? What's implied but not stated? What would a sceptical reader push back on here? The answers don't replace my revision process. They just make it more informed and, honestly, a little less lonely.

Rewrite the Output Until It Feels Lived

Even when AI gives me something genuinely useful, I almost never keep the language as-is. I rewrite until the sentences feel like mine — which usually means adding a real observation somewhere, removing inflated certainty, or simplifying a phrase that sounds more impressive than true. If a line reads smoothly but could have been written by anyone, it usually needs more work. I want the finished piece to carry evidence that a person stood behind it and made choices.

That final human pass is where authorship becomes visible again. Readers may not know exactly how a piece was drafted, but they can feel when writing contains experience, limits, and earned specificity. AI can help me get closer to that standard — but only if I insist on finishing the work as a person, not a curator of machine output.

Used carelessly, AI can produce the illusion of thought. Used well, it can make your thinking more demanding and more reflective. The difference comes down to what questions you ask, when you ask them, and whether you're willing to stay responsible for the final meaning. That's the version of AI I want in my workflow — not a ghostwriter, but a thinking partner that pushes back.