AI as Critic: Prompts That Earn Real Pushback

By Better Wise

You have already built an argument from claims, reasons, and assumptions, and you have stress-tested the sources behind your evidence. That work puts you in the right position for this lesson: you now have something worth attacking. The risk from here is subtle. A generic "what do you think of my draft?" invites the AI to be agreeable, vague, or worse, to hand you a cleaner paragraph that quietly replaces your thinking. This lesson teaches you to write critique prompts that make the AI attack your argument from a named perspective, while leaving the writing to you.

By the end of this lesson you will be able to:

• Explain why a vague feedback request produces weak, unusable pushback.
• Build a critique prompt with five working parts.
• Assign named counter-perspectives that attack different weak points.
• Sort the AI's objections into fix, verify, and set aside.
• Ask a second round of questions that surfaces the objection it held back.

Why a Friendly "What Do You Think?" Fails

Compare two requests about the same paragraph on the project's sample argument, the claim that the city library should drop overdue fines.

Weak: "Here is my paragraph. What do you think?"

Stronger: "Argue against this claim as a skeptical budget officer. Do not rewrite it. List the three strongest objections, ranked by how much damage they do."

The weak request gives the model no role, no target, and no output shape, so it falls back on its default habits: summarizing your point back to you, complimenting the structure, and offering a polished rewrite. None of that tells you whether the argument survives contact with an opponent. The stronger request installs an opponent.

One more thing to know before you start: an AI critic is easy to satisfy and easy to fool. It will produce confident objections and confident "evidence" for them. Your job is to harvest the objections and then check anything that sounds like a fact, which is exactly the verification habit Module 3 opens with.

The Five Parts of a Prompt That Earns Pushback

A critique prompt that reliably produces usable material has five parts.

A reusable template:

Perspective: You are a skeptical [specific role] who believes my claim is wrong.
My claim: ...
My reasons: 1) ... 2) ... 3) ...
My evidence: ...
Your job: name the three strongest objections to this argument.
For each: state the unstated assumption it attacks and the evidence that would answer it.
Constraints: Do not rewrite my text. Do not offer replacement sentences. Do not praise.
Output: numbered list, ranked from most to least damaging.

The constraint line is the one beginners skip, and it matters most. Without it, the model's helpfulness instinct turns your critique session into a ghostwriting session.

Try it (project thread). Here is a weak request about your capstone argument: "I wrote three reasons why the library should drop fines. Can you give me feedback?" Rewrite it so it contains all five parts.

Click to reveal guidance

Your version should name a lens (a hostile resident, a city finance analyst, a librarian who enforces fines daily), paste your actual claim, reasons, and evidence, ask for objections to your assumptions rather than general comments, forbid rewriting, and specify a ranked numbered list. If your rewrite still says "feedback," you have not named the attack yet.

Lenses Worth Assigning and What Each One Hits

Different lenses attack different parts of an argument. Rotating through two or three is usually enough for one draft.

flowchart TD
  Draft["✍️ Your draft"] --> Prompt["🎯 Critique prompt with a named lens"]
  Prompt --> Obj["⚔️ Numbered objections"]
  Obj --> Sort{"Sort each one"}
  Sort -->|must fix| Revise["You rewrite the weak link"]
  Sort -->|verify| Check["Check the source or data"]
  Sort -->|set aside| Note["Note it, move on"]
  Check --> Revise
  Revise --> Draft2["🔁 Next draft, still yours"]

• The hostile reader. "Assume my claim is false and build the best case against it." This produces the counterargument you will need anyway.
• The steelman builder. "Give me the strongest, most charitable version of the opposing view, not a weak one." Weak opponents make your argument look better than it is.
• The methodologist. "Examine whether my evidence actually supports my claim, including sample size, source, and alternative explanations." This one audits the work you did on stress-testing sources.
• The practitioner. "You run a small business, a classroom, or a department budget. Where does my proposal break in practice?" Good for feasibility and cost objections.
• The confused newcomer. "You are reading about this topic for the first time. Point to the sentence where you stop following the reasoning." This finds buried assumptions and undefined terms.

Try it. Choose two lenses for your capstone argument, write the single sentence that assigns each, and predict which part of your argument each will attack.

Click to reveal guidance

A strong prediction might read: the budget officer will attack my cost estimate, and the steelman builder will attack my claim that fines do not change borrowing behavior. If you cannot predict what a lens will hit, you may not understand your own weak points well enough yet. Write down the prediction before you send the prompt, then compare.

Sorting the Pushback: Fix, Verify, Set Aside

Raw objections are not instructions. Read them, then sort each one into three piles.

1. Must fix. The objection names a gap you can act on right now: a missing reason, an unstated assumption, a leap in logic. Example: "Your first reason assumes that people who stop borrowing do so because of fines, but your source only shows that both numbers moved together." That is a structural weakness in your argument, and no source can rescue it.
2. Verify. The objection depends on a factual claim you have not checked, or on the model's memory of a study, number, or quote. Example: "The 2019 survey you cite had only forty respondents." Possibly true, possibly invented. Check the source yourself before you rewrite a word.
3. Set aside. The objection is real but off-target for this draft, or it concerns style when you asked about argument. Example: "The phrase 'ordinary families' is vague." Worth noting for revision, not worth restructuring the argument.

Two rules keep this honest. First, never delete an objection because it stings; sort it by what it asks you to do. Second, never accept an objection because it sounds expert. A confident tone is not evidence.

Try it. Sort these three objections about the library argument:

1. "You assume residents who stop using the library do so because of fines; your evidence only shows a correlation."
2. "The study you cite surveyed forty people, which may not support a citywide claim."
3. "Some readers may find 'ordinary families' vague."

Click to reveal guidance

1 goes to must fix: it exposes a causal assumption your reasons depend on, and you need a reason or a hedge in the claim, not a new source. 2 goes to verify: look up the study yourself, confirm the sample size, and decide whether it still carries the weight you gave it. 3 goes to set aside for now, unless the vagueness is load-bearing for your claim, in which case it becomes a small fix.

Round Two: Asking for the Objection It Held Back

One pass of critique rarely reaches the deepest problem. After you have sorted the first list, send one follow-up that forces the model past politeness:

• "What is the strongest objection you did not raise, and why did you leave it out?"
• "Which of these objections would an expert in this field call fatal, and which would they call a nitpick?"
• "What evidence would make you abandon my claim entirely?"

That last question is the most useful one in the lesson. If you cannot name what would change your mind, your claim may be a belief wearing an argument's clothes.

Rules of Thumb for Critique That Stays Yours

• Finish your own draft before you open the chat. Critique of a blank page becomes generation.
• Paste your actual text. The model cannot attack what you only describe.
• Always include the no-rewrite constraint, and mean it. If a polished paragraph appears anyway, do not paste it into your document. Read it, extract the objection it implies, and write the revision in your own words.
• Ask for objections ranked by damage, not by order of appearance. You have limited revision time.
• Send your predictions first, then compare. Being wrong about your own weak point is the most valuable information the session produces.
• Treat every factual claim in an objection as unverified until you check it yourself.

What to Carry Into the Checkpoint

• A weak request gets praise or a rewrite. A prompt with a named perspective, your pasted text, a specific job, a no-rewrite constraint, and a fixed output format gets objections you can use.
• Different lenses hit different weak points, so rotate two or three rather than asking for "feedback."
• Sort pushback into must fix, verify, and set aside, and verify anything that smells like evidence.
• The goal is not a better paragraph from the AI. It is a sharper list of problems for you to solve.

Next up is the Module 2 checkpoint, where you will assemble your argument and run a full critique round on it, from prompt construction to sorted objection list.