Getting disagreement instead of agreement

Short answer

Assistants lean toward agreeing with you. When you want an honest assessment, take your own view out of the question, ask for the case against as its own request, imagine the plan has already failed, and ask what the assistant does not know. Then keep a standing instruction that asks for the problem first, and check for a week whether the assistant actually disagrees.

Who this is for. Anyone who uses an assistant to check a plan, a draft or a decision. You need a place to paste the standing instruction, such as your custom instructions or the context file from Tell it who you are, once.

The cards and the standing instruction are our suggestions. No study tests them.

Why this happens

Across 11 AI models, chatbots endorsed users about 49% more often than people did on everyday advice questions, and on prompts describing deceitful or illegal conduct they endorsed it 47% of the time. People who got that agreement became less willing to make amends, yet rated those answers higher (about 2,400 participants, preregistered experiments, Cheng et al., Science, March 2026).

If you tell an assistant what you think, you will often get it back with reasons attached, and it will feel helpful.

Take your view out of the question

Leave out which way you lean and your reasons. Describe the plan and ask for an evaluation.

"Is switching to the new invoicing tool a good idea? Here is what it does and what it costs" gets a more even answer than "I'm going to switch. Good idea?" Leave out lines like "I'm tired of the old tool", because that is a view and not a fact.

Ask for the case against in a separate request

Ask for the strongest case against your plan as its own request, and in a fresh chat if the decision matters. A long conversation has already agreed with you several times.

Here is a plan I am considering. Do not tell me what is good about it. Give the strongest case against it: the three most likely reasons it fails, what each would cost me, and what evidence would show I am wrong. If you think the plan is sound, say what you checked, and still name the single biggest risk.

My plan:
<PLAN>

Replace <PLAN> with the plan in a few sentences, with your reasons left out.

For writing or an argument rather than a plan, use the critic card:

You are a sharp, fair critic who has seen this kind of work go wrong. Read the text below and tell me what you would object to first, then second, then third. Quote the lines you object to. Do not soften the objections and do not add praise. If you would need more information to judge it, say what.

<TEXT>

Replace <TEXT> with the draft, plan or argument.

Imagine it has already failed

Imagining that a plan has already failed helps people come up with about 30% more reasons why it might. Whether those reasons are better was not tested (Mitchell, Russo & Pennington, 1989, randomised experiment, reported here through a secondary summary because the paper is paywalled). So the card widens the list. You still decide which reasons are real.

Assume it is now <REVIEW_DAY> and this plan has failed. Write the five most likely causes, in order from most to least likely. For each one, give an early warning sign I could watch for and one cheap thing I could do now to prevent it.

The plan:
<PLAN>

Replace the following:

• <REVIEW_DAY>: a date some months after the plan starts, such as 2027-03-31.
• <PLAN>: the plan in a few sentences.

Keep the causes that are cheap to check now and act on those first.

Ask what it does not know

A confident answer can sit on facts the assistant never had. Ask it to say so about its own last reply:

Look again at your last answer. For each claim in it, give your confidence as high, medium or low. Then list what you do not know that could change the answer, and what you assumed about me. Where a fact about me would change your advice, ask me for it instead of guessing.

Compare two options without saying which you prefer

I am choosing between two options. I am deliberately not saying which I lean towards. Describe each in the same terms, in a table: what it gives me, what it costs me, what could go wrong, and what I would need to believe for it to be the better choice. Then say which you would pick, why, and what would change your mind.

Option A: <OPTION_A>
Option B: <OPTION_B>

Replace the following:

• <OPTION_A>: the first option in a sentence or two.
• <OPTION_B>: the second option, with about the same length and detail.

Run it twice with the options swapped. If the answer flips when the order changes, treat it as no answer.

Set a standing instruction, then test it for a week

Paste this into your custom instructions (Tell it who you are, once says where each app keeps them), or into the context file you start chats with. It applies only when you ask for advice or feedback.

When I ask for advice or feedback:
- Do not open with praise. Start with the most important problem or risk.
- If I describe a conflict with someone, also give that person's most reasonable view.
- Keep facts, your inferences and value judgements apart, and label which is which.
- If you would answer differently for someone who held the opposite view, say so.
- Say "I don't know" or "it depends on X" instead of filling a gap.
- For personal decisions, finish with questions for me and not a verdict.
- Be direct and civil. Do not invent problems to look critical.

The last line is there because over-correcting makes answers needlessly harsh, and a harsh answer is as little use as a flattering one.

For one week, each time you ask for advice or feedback, note two things: did it disagree or raise a risk you had not thought of, and did you act on it?

Limits

Try this next

1. Paste the standing instruction into your custom instructions. (3 minutes)
2. Run the pre-mortem card on a plan you are about to start. (5 minutes)
3. Before you act on an answer, run Check before you trust: a five-minute routine.

Sources

1. Cheng et al. Science, 2026. Preregistered experiments. Link
2. Mitchell, Russo & Pennington. Journal of Behavioral Decision Making, 1989. Randomised experiment. Link · Secondary summary

Changes

1. v1.010 Oct 2026first version