You are writing an essay. Maybe a personal one about the year you spent driving your mother to dialysis, maybe an argument about why your profession is wrong about something, maybe an op-ed you have been putting off for a month. You have a feeling and a pile of notes. What you do not yet have is the claim, and you suspect the assistant will be happy to supply one, which is exactly the problem.
Nonfiction with AI fails in two directions. In one, the model supplies the argument, and it is the median argument: reasonable, balanced, forgettable, with a study attached that may not exist. In the other, the writer supplies the argument and the model sands it down, hedging each claim and adding a paragraph for the other side until the essay has no spine. Both failures produce something that reads fine and says nothing a reader could disagree with.
The method here is the same loop as everywhere on this site, with the weights shifted. You direct harder: the claim, the sources, the memories, and the stance are yours before the chat opens. The assistant proposes structure, transitions, objections worth answering, and a list of what needs checking. You judge every hedge and every invented specific. And you own the facts, which in nonfiction means you opened the source yourself.
What changes for this format
The method is the same everywhere. These are the places where this format bends it.
The argument has to exist before the chat does
An essay is a person thinking in public, and the model does not have a position. If you ask it what to argue, it will give you the view that offends the fewest readers, which is also the view nobody needed an essay to hear. Before you draft, you should be able to say in one sentence what you believe and in one more sentence why a reasonable person disagrees. If you cannot, you are not ready to draft, and that is fine. Use the Discover Writing Ideas prompt to find the tension you keep circling. Use it to find your claim, never to be handed one.
Sourcing: the model will cite what it does not have
Ask for evidence and the assistant will produce it: a 2019 study, a 34 percent figure, a quote from a named researcher, all shaped exactly like the real thing. Some will be real. Some will be a plausible blend of real things. You cannot tell by reading. The rule for this path is that every fact in the essay traces to something you opened yourself, with the link or page number in your source note. The brief forbids invention and requires [VERIFY] tags, and the Fact-Check Pass makes the assistant list every claim so you can check each one. Treat it as a research assistant that tells you what to look up, never as the thing you looked up.
The both-sides reflex
Ask the model to argue and it hedges. While some contend X, others believe Y, and ultimately the truth lies somewhere in between. Ask it to add a counterargument and it writes one so generous that your own position looks unkind beside it. This is a default rather than a judgment: the model is trained to be agreeable to every reader at once. Direct it. Your stance goes into the accepted-decisions ledger as a protected item, and the instruction is explicit: nothing hedges the claim, the objection gets one paragraph stated as strongly as its best advocate would state it, and the next paragraph answers it. If the answer is not good enough, that is information about your argument and a reason to think harder, never a reason to soften.
Personal material the model cannot know
In a personal essay the raw material is your life: the hallway outside the dialysis unit, the specific wrong thing you said in the car, your father's handwriting on the envelope. Leave a gap and the model fills it with plausible texture: a worn leather jacket, a faded photograph, sunlight through the curtains. Those are not your memories, and a reader can feel the difference even when they cannot name it. Write the memory passages yourself, however rough, and hand them over marked mine. Let the assistant work on transitions, order, and the paragraphs of argument around them. It is a good editor of a memory and a bad inventor of one.
Structure follows the turn in your thinking
Left alone, the model produces the five-paragraph shape: introduction with thesis, supporting sections of equal weight, conclusion that restates. An essay moves differently. It has a turn: I believed this, then this happened, now I think this instead and here is what it costs me to think it. Give the assistant the turn as the outline's backbone and reject any proposal that comes back as a list of parallel points. The section where you change your mind should be longer than the others, because it is where the essay is.
The ending the model wants to write
It wants to tell the reader what it all meant. Ultimately, I learned that. In the end, what matters is. The lesson paragraph is so reliable you can plan for it: it will be there, and you will cut it. The better endings in nonfiction are usually one step before the meaning, an image or a fact that lets the reader arrive on their own. Ask the assistant to stop the draft before the final paragraph, then write the last lines by hand. If you want the model's help there, ask for five different last sentences and expect to use none of them as written.
The process
- 01You and AI
Find the claim by talking it out
Open a chat with no intention of drafting. Dump what you are circling: the notes, the anecdote you keep telling, the thing that annoyed you enough to write. Ask the assistant to reflect back the tensions it hears and to ask you questions, rather than to propose a thesis. Somewhere in your answers is the claim. When you can say it in one sentence and name who disagrees, write both down. That pair is the first entry in your ledger.
Send this I am trying to find the argument of an essay, not write it. Here is everything I have: [NOTES AND RAW MATERIAL]. Do not propose a thesis. Instead, list the tensions you hear in my notes, in my words where you can quote them, and ask me up to five questions that would help me decide what I actually think. One list, then wait.
- 02You
Write the evidence and the memories by hand
For an argumentative essay, build the source note: every fact you intend to use, with where it came from and a line about how confident you are in it. For a personal essay, write the scenes. Not well, just true: what was said, what was in the room, what you did next. These blocks are the raw material. Everything the assistant drafts has to be built from them, and anything it needs that is not there becomes a [VERIFY] tag, never a paragraph.
- 03AI proposes
Outline from the turn, not the template
Use the Rough Idea to Outline prompt with the claim, the disagreement, and the turn: what you believed, what changed, what you think now. Ask for an outline whose sections are claims the reader could argue with, uneven in length, with the objection placed where it does the most damage rather than tucked near the end. Reject the first outline if it is a list of parallel points. That is the five-paragraph shape wearing a different coat.
- 04AI proposes
Brief the session and draft section by section
Use the Personal Essay mega prompt for personal material, or the universal brief for argument. Paste the memory blocks marked mine and the source note as the only source of facts. Ask for one section per turn, in order. Expect the argument sections to come back hedged and the memory sections to come back decorated. Both are predictable, and both are bounded edits rather than rewrites.
- 05You
Hunt the hedge
Read every claim and ask whether it is stated at the strength you believe it. Softened claims are the model's most invisible edit: a somewhat, an often, a perhaps, a many would argue. Restore them by hand or with a bounded request. Then read the counterargument paragraph and check that your answer to it is actually an answer. If it is not, the essay has found a weak point in your thinking, and the right fix is to think, not to prompt.
Send this Change only the counterargument paragraph. The problem: it concedes so much that my claim looks unreasonable beside it. State the objection once, as strongly as its best advocate would, in no more than three sentences. Then answer it using only the facts in my source note. Do not soften my claim anywhere else. Show OLD and NEW, then wait.
- 06You
Fact-check with the sources open
Run the Fact-Check Pass so the assistant lists every factual claim, every number, every attributed statement, and every place it could not trace a claim to your source note. Then do the checking yourself with the sources open in another window. The assistant cannot verify anything, and it will say so if your brief is right. If it ever reports that it has confirmed a fact, treat that as a failure of the brief and a reason to check twice.
- 07You
Cut the lesson, then read it aloud
Delete the final paragraph the model wrote. Write the last lines yourself, aiming one step short of the meaning. Read the whole essay aloud and rewrite by hand any sentence that sounds like a different person. If the essay is for a class, a publication, or a contest, read their policy on AI assistance and follow it exactly. Where the policy is silent, a short note about what you used it for is the honest default.
Toolkit
The prompts that fit this format, and when to reach for each.
- Mega promptPersonal Essay mega promptFor essays built on your own experience. It keeps your memory blocks untouched and puts the assistant to work on order and transitions.
- Stage promptDiscover Writing IdeasBefore you have a claim. Use it to find the tension you keep circling, never to be handed a thesis.
- Stage promptRough Idea to OutlineOnce you can state the claim and the disagreement. Give it the turn in your thinking and reject a parallel-points outline.
- Stage promptFact-Check PassAfter the full draft, before anyone else reads it. It lists the claims, and you check them with the sources open.
Before and after
A paragraph the way an assistant tends to return it, and the same paragraph after a writer took it back.
Teaching my father to use his new phone was, in many ways, a journey for both of us. As I watched his weathered hands struggle with the tiny icons, I was reminded of all the times he had patiently taught me things when I was young. There is a profound role reversal that happens as our parents age, and it forces us to confront our own mortality. In the end, I realized that the phone was never really about the phone. It was about connection, patience, and the enduring bond between a parent and a child.
My father holds the phone the way he holds a hammer, which is the problem. He is seventy-four and has held a hammer for sixty of those years, and the phone keeps reading his grip as a request to open the camera. I had not planned to spend Thanksgiving on this. I had planned to spend it not talking about my sister. By four o'clock he could send a text, if the text was ok, and he had sent it to me eleven times from across the table, looking up each time to see whether it had worked. It had. I did not tell him he could stop.
The invented texture went first. Weathered hands and tiny icons could belong to anyone's father. The hammer grip, the camera, and the sister are things only this writer knew, and they came from her memory block rather than the chat. The argument moved out of the paragraph and into what happens: role reversal and mortality were the model's thesis, and the writer's real subject, what she was avoiding, now arrives sideways through the sister and the eleven texts. And the ending refuses the lesson. The enduring bond is gone. In its place is an action the reader can interpret without help.
Pitfalls
The thesis is reasonable, balanced, and you do not actually believe it.
The model supplied it. Go back to Discover and talk until you can say what you think in one sentence and name who disagrees. Then put that sentence in the ledger as protected and brief the draft from it.
A study or statistic appears that you did not provide.
Delete it. Ask the assistant to list every fact in the draft and where each came from in your source note. Anything untraceable becomes [VERIFY] or goes. Never let the assistant resolve its own [VERIFY] tags.
A memory in the personal essay has details you do not remember.
The model decorated a gap. Replace the passage with your own rough version marked mine and instruct the assistant to edit it only for the problem you name, never for texture.
Every claim has grown a somewhat or a perhaps.
Run a dedicated hedge pass: ask for a list of every qualifier the draft added that was not in your notes, then restore the claims by hand to the strength you hold them.
The outline is five parallel sections of equal length.
Give the assistant the turn in your thinking as the backbone and ask for uneven sections. If it cannot produce an outline that is not a list, write the outline yourself in six lines and ask only for transitions.
The last paragraph explains what the essay meant.
Cut it. Write the final lines yourself, stopping one step before the meaning. If you want options, ask for five different last sentences and treat them as prompts for your own.
Questions
Is it okay to use AI to write an essay for school?
That depends entirely on your instructor's or institution's policy, and you should read it before you open a chat. Where AI is allowed, the honest use is the one described here: your claim, your sources, your memories, with the assistant on structure and critique, and a disclosure of what you used it for. Submitting a model's argument as your own is a different act, and most policies name it.
How do I keep the argument mine when the model keeps softening it?
Put the claim in your ledger as a protected decision in the exact words you believe it, and tell the assistant that nothing hedges it. After each turn, read for added qualifiers. When the model keeps reintroducing balance, that is usually a sign to reset the chat with the ledger rather than to argue with it.
Can the assistant find sources for me?
It can tell you what kinds of sources would support a claim and what terms to search, which is useful. It cannot be trusted to produce the sources themselves. It will generate citations that look right and may not exist. Treat every reference it offers as a search query, open the real thing, and cite only what you have read.
Which parts of a personal essay should I write by hand?
The memories, in whatever rough form you can manage, and the ending. The model can order them, write the connective paragraphs, and ask the questions that make a scene sharper. It cannot know what the room smelled like or what you said in the car, and when it guesses, readers feel it.