Isn't the model just making up plausible arguments?

Partly, which is why step 3 exists. Treat the steelman as a list of hypotheses worth checking, not as evidence — the checkable items are the output that matters.

Can I use this for exam essay questions?

Yes, and it is one of the best uses of it — provided you write the essay. The pattern surfaces the counter-argument you would have missed; using its prose as your answer skips the only part that teaches you anything.

Docs & research

Steelman then gaps

Get the best version of an argument you disagree with, then the specific evidence that would change your mind — instead of a balanced-sounding list of nothing.

Use this before a decision you are already leaning on, or when you need to argue a side you did not pick.

The pattern

Copy this and replace everything in [SQUARE BRACKETS] with your own specifics.

Prompt template

Position: [THE CLAIM, stated as sharply as you can.]

Step 1 — Steelman it. Give the strongest honest case FOR this position: [3-4] arguments, each with the reasoning made explicit and the strongest piece of evidence you know of. No strawmen, no hedging, no "some people say".

Step 2 — Now attack it. For each argument, name the specific weakness: an assumption that may not hold, evidence that would contradict it, or a case where it fails.

Step 3 — What would change my mind? List the [2-3] concrete observations or facts that, if true, should make me abandon this position. Make them checkable.

Do not conclude. Do not tell me the truth is somewhere in the middle.

Why each part is there

Stated as sharply as you can: A vague position cannot be tested. “AI will change education” has no steelman; “schools should stop setting take-home essays” does.
Steelman before attack: Order matters. Asked for pros and cons together, the model produces a symmetrical, weightless list. Building the strongest case first forces it to actually inhabit the position.
What would change my mind: The part that makes this useful rather than interesting. It converts an opinion into something falsifiable that you can go and check this week.
Do not conclude: Blocks the reflexive “ultimately, balance is key” ending that quietly cancels everything above it and leaves you exactly where you started.

Before and after

The lazy version

What are the pros and cons of students using AI for homework?

What goes wrong: Six pros, six cons, all about equally weighted, all uncontroversial. Nothing in it could ever change a decision, because nothing in it is committed to anything.

The pattern, filled in

Worked example

Position: A Class 9 student learning AI should spend the first six months on mathematics and Python before touching any machine-learning library.

Step 1 — Steelman it. Give the strongest honest case FOR this position: 4 arguments, each with the reasoning made explicit and the strongest evidence you know of. No strawmen, no hedging.

Step 2 — Now attack it. For each argument, name the specific weakness: an assumption that may not hold, evidence that would contradict it, or a case where it fails.

Step 3 — What would change my mind? List 3 concrete observations that, if true, should make me abandon this position. Make them checkable.

Do not conclude. Do not tell me the truth is somewhere in the middle.

What that buys you: Four real arguments, four real weaknesses — including the strongest one, that early abstraction kills motivation — and three checkable observations, such as “learners who start with libraries show equal or better maths retention at month 12”. That is a research question, not a debate.

Now do it on your own work

Pick a position you actually hold about your own studies or work. Run all three steps. Then go and check one of the “would change my mind” items against reality — ask someone, look it up, run the experiment.

Check yourself — every box should be true

Your position statement is sharp enough that a reasonable person could disagree with it.
The steelman contains at least one argument you had not thought of.
Every “would change my mind” item is checkable — you can say who or what you would ask.
You checked one of them, and recorded what you found.

Then explain it back

One argument in the steelman was stronger than you expected. Which one, and has it changed how confident you are — or only how you would defend your view?

Write two or three sentences in your own words. If you cannot, you copied a prompt rather than learning a pattern — which is the whole difference this library is trying to make.

Does this depend on which AI tool I use?

Alignment tuning matters here: some assistants soften step 2 or refuse to argue sharply for a contested position, and add a balanced conclusion regardless of your instruction. If yours does, split it into two separate prompts — one that only steelmans, one that only attacks — so neither can hedge against the other.

Why it actually works

Naming in advance what would falsify your belief is the discipline behind a held-out test set: you decide what would count as being wrong before you look.

Why we teach the reasoning, not the tool →

Prompting well is a skill you can pick up in a week. Understanding what is happening underneath is the 18-month curriculum — and it is what stays useful when the tools change.

FAQ

Isn't the model just making up plausible arguments?: Partly, which is why step 3 exists. Treat the steelman as a list of hypotheses worth checking, not as evidence — the checkable items are the output that matters.
Can I use this for exam essay questions?: Yes, and it is one of the best uses of it — provided you write the essay. The pattern surfaces the counter-argument you would have missed; using its prose as your answer skips the only part that teaches you anything.

Next patterns

The Prompt Patterns Library is a separate, optional track — it does not change the 78-week curriculum or your roadmap. See what the curriculum actually covers →