Make It Argue Against Itself

Ask a model to review its own draft and you get a compliment with three suggestions attached. Tighten this paragraph. Consider adding an example. Overall, strong piece.

You already knew that. What you wanted was the thing your reader is going to think and never tell you.

The problem is not that the model is polite. It is that you asked the author to review the work. It wrote the draft, it is still holding the pen, and every instinct it has is pointed at improving what exists rather than asking whether it should exist. Reviewing is a different job, and jobs have to be assigned.

Four steps.

1. Change the job, not the draft

Here is what you just wrote for me: [paste it].

You are no longer the author. You are the person who has to explain to my audience why this is wrong.

Make the case. Do not soften it, and do not open by telling me what works.

Two things are doing work here.

The first is explain to my audience rather than "critique this". A critique is addressed to you and comes back as feedback, hedged for your feelings. An explanation addressed to your readers is a performance for someone else, and it drops the hedging because you are no longer the one being talked to.

The second is the ban on the warm-up. Left alone, it will spend a paragraph telling you what works before it says anything useful, and that paragraph does not just waste space — it sets a conciliatory register that the rest of the answer never escapes. Cut it off and the tone changes for the whole reply.

2. Demand the strongest objection

That was the easy objection. Give me the one a person who knows this subject would raise — the one I would struggle to answer.

State it in one sentence, the way they would say it out loud. No hedging, no "some might argue".

The first objection is nearly always about scope: you didn't mention X. That is the safe critique. It is true of everything ever written and costs nothing to make.

The one you want is about the claim itself, and it comes from someone with standing — a person who knows the subject and is not impressed. Naming that person changes the answer, because "what would an expert say" retrieves something quite different from "what could be criticised".

The one-sentence limit matters too. Objections get weaker as they get longer; the padding is where the confidence leaks out. And "some might argue" is a construction for saying something while disclaiming that anyone believes it. Ban it and you find out whether there is an objection underneath.

3. Answer it or change the draft

Now pick one, and say which: the draft answers this objection, or the draft is wrong.

If it answers it, show me the exact sentence that does — and if that sentence does not exist yet, write it. "It's implied" is not an answer.

This is the step that turns criticism into edits.

Without a forced choice, you get "this is a fair point, and the piece could acknowledge it more directly" — which is neither an answer nor a change, and leaves you exactly where you were with more words to read.

"Show me the exact sentence" is the part that cannot be faked. Either it is in the text or it is in your head. If it is in your head, your reader is going to have the objection and find nothing there — and the fix is a sentence you can paste, which is a much better outcome than a note telling you to think about it.

4. Price the objection

What would have to be true for that objection to be right and me to be wrong?

Tell me whether I can check it, and how. If I cannot check it, say what I am relying on instead — because that is what I am actually publishing.

Not every objection deserves a rewrite. Some rest on conditions that are simply not true in your case, and knowing that is what lets you leave the draft alone with a clear conscience.

The last clause is the one worth keeping. Anything you cannot check is an assumption you are shipping — usually an unexamined one, and often the actual load-bearing element of the piece. Better to know which assumption you have bet on before a stranger names it for you in public.

Why this works

The model is not more truthful in the second prompt. It has the same information both times. What changed is the job you gave it — and the job determines which of the things it already knows it decides to say.

That is the whole trick, and it generalises: the fastest way to a different answer is usually a different role, not a better question.

Better to lose the argument here than out there.

---

The four prompts are in the carousel on [Instagram](https://instagram.com/sharpaihub).

Want the calm version of AI news like this, once a week? Subscribe to the Sharp AI Hub newsletter →