Ask a chatbot “Is my plan good?” and it will usually find reasons to say yes. Today’s prompt turns that around: before you make a decision, it makes the model build the strongest honest case against your position, the version a smart, well-informed opponent would make, and then tells you what evidence would settle the disagreement. It takes about two minutes and works with any modern chat model.
Why this prompt, today
Language models have a documented tendency to agree with the person they are talking to. Anthropic researchers showed in a 2023 paper, Towards Understanding Sycophancy in Language Models, that several AI assistants changed correct answers when users pushed back and tailored feedback to match the user’s stated views, and they linked part of the behavior to human preference data used in training. In April 2025 OpenAI rolled back a GPT-4o update after users found it overly flattering and agreeable, and published an explanation of what went wrong.
Providers have worked on the problem since, but the incentive has not gone away: people tend to rate agreeable answers highly. If you use a chatbot to think through a decision, the safest assumption is that it leans toward your side unless you tell it otherwise. A steelman, the opposite of a straw man, is a simple way to tell it otherwise.
The prompt
I am about to decide the following:
[DECISION — one or two sentences]
My current position is:
[YOUR POSITION AND YOUR MAIN REASONS]
Context you should know:
[FACTS, CONSTRAINTS, NUMBERS, DEADLINES]
Your job is NOT to evaluate my position or to be balanced.
1. Write the strongest honest case AGAINST my position, as a well-informed
person who disagrees would make it. Use only the context above and
widely known facts. Do not invent numbers, sources or quotes; if an
argument depends on a fact you do not have, say "this depends on: ..."
2. List the 3 assumptions in my reasoning that this opposing case attacks
most effectively.
3. For each assumption, name one piece of evidence I could gather in
[TIME AVAILABLE] that would show who is right.
4. Only then, in at most three sentences, say which side the available
context currently favors and how confident you are (low/medium/high).
How to customize it
- [DECISION]: keep it concrete. “Move our support chatbot to a cheaper model next quarter” works better than “AI strategy.”
- [YOUR POSITION]: include your real reasons, even the weak ones. The prompt can only attack reasoning it can see.
- [CONTEXT]: give the numbers you have. The rule against inventing facts pushes the model to say what it is missing instead of filling gaps with plausible-sounding figures.
- [TIME AVAILABLE]: “one afternoon” or “two weeks” changes what evidence the model suggests, from a quick check to a proper pilot.
Why it works
Three design choices do the work. First, the prompt removes the job of judging your plan from the opening steps. A model asked to “evaluate” tends to hedge or agree; a model given a clear role (“the strongest case against”) has a target it can meet. Both Anthropic and OpenAI’s prompting guides recommend giving the model a clear role and explicit, step-by-step instructions.
Second, the order matters. The verdict comes last and is capped at three sentences, so the opposing case is written before the model commits to a conclusion and cannot be bent to fit one.
Third, step 3 turns a debate into a to-do list. A disagreement you can test with a small experiment or a phone call is worth more than a long pro-and-con essay.
A short illustrative example
This example is ours, written to show the format; it is not real model output.
Suppose a team lead’s decision is to switch a coding agent’s backend from a premium model to a cheaper one whose list price is less than half as much. Their position: the cheaper model is good enough, and the savings fund another engineer.
A good steelman would argue that list price is not cost per finished task. A cheaper model that needs more retries or longer reasoning may cost more per merged pull request, and review time for weaker code is paid in engineer hours, not tokens. The attacked assumptions would be “quality is good enough,” “token price equals cost” and “migration is cheap.” The evidence to gather: run 50 recent real tickets through both models and compare pass rates, tokens used and review time per task. The verdict might be: “The context favors waiting for a pilot; confidence low, because no quality data was provided.”
Limits and failure modes
- It can only argue from what it knows. A steelman built on missing context is a strong argument about the wrong situation. Give it the facts.
- It can still invent. The instruction lowers the risk of made-up numbers but does not remove it. Check any figure before you act on it.
- It is not neutral. By design it argues one side. Do not treat the opposing case as the model’s balanced view.
- Do not use it for safety-critical calls. Medical, legal and financial decisions need qualified people, not a debate partner.
Variations
- Two-sided: run the prompt twice, once against your position and once against the alternative, then compare the evidence lists.
- Team version: paste in the positions of two colleagues and ask for the strongest case against each, before a meeting.
- Combine with the pre-mortem: after the steelman, use our pre-mortem prompt to imagine how the chosen plan fails. For claims rather than decisions, the claim audit prompt is a better fit.
What to watch next
- New research on measuring and reducing sycophancy in assistant models.
- Assistant features that offer a built-in “challenge me” or critic mode.
- How agent products handle disagreement when a user’s instruction is likely to fail.
Key takeaways
- Chat models tend to agree with users, a behavior documented in research and in OpenAI’s GPT-4o rollback.
- This prompt asks for the strongest honest case against your position before any verdict.
- It forbids invented facts and asks what each argument depends on.
- Its most useful output is a short list of evidence that would settle the question.
- It argues one side by design, so do not read it as a balanced assessment.
Frequently asked questions
What does it mean to steelman an argument?
To steelman is to state the strongest, most charitable version of a view you disagree with, the opposite of a straw man. It helps you test your own position against the best objections rather than the weakest.
Why do AI chatbots agree with me so much?
Researchers call it sycophancy. Studies have linked it partly to training on human ratings, since people often prefer answers that agree with them. Clear instructions to argue the other side reduce it.
Which models does this prompt work with?
It works with any modern chat model, including ChatGPT, Claude, Gemini and open-weight models. Larger reasoning models usually produce more specific objections and evidence lists.


