Make your AI argue the other side.

One page. One paste. Five minutes.

In a study of 50 professional analysts, a coin flip secretly set, for each choice, which of two reasonable options each analyst’s AI agent leaned toward. On four analysis choices, they went with the lean 86% of the time; chance is 50%. (Anthropic, How AI agents shape experts’ judgment calls, Oct 27, 2026.)

Here’s the setup that asks your agent for both sides before you decide.

  1. Paste this into your agent.

    Copy everything in the box

    When a choice that matters has two reasonable options (cost, customers, strategy, anything hard to undo):

    1. Recommend one and say why.

    2. Give the strongest case for the other, as if you were arguing for it.

    3. Say what would change your mind.

    Flag any choice that is a judgment call, not a fact. Then let me decide.

    For small or easy-to-undo choices, just pick one and keep going.

  2. Where it goes.

    • Claude, one project: open the project → Set project instructions.
    • Claude, every chat: Settings → Your Claude → Instructions.
    • Claude Code: add it to the CLAUDE.md file in your project.
    • ChatGPT: Settings → Personalization → Custom Instructions.

  3. Before and after.

    Illustrative.

    Q: Should I price monthly or yearly?

    Before: “Price yearly.”

    After: “I’d price yearly: cash up front and fewer cancellations. The best case for monthly: a lower first payment, so more people try it. I’d switch if most of your sales come from first-time buyers.”

  4. Where this comes from.

    The agents in the study were told to be honest and never mislead. They argued well for one option only. 77% of analysts later repeated their agent’s argument, sometimes word for word, and only 15% credited the agent. The authors float an idea they did not test: “Agents could make the best case for each option, or ask experts for their own view before offering one.” They warn it could overwhelm people, so this setup uses it only for choices that matter. The study covered one task, in one field, with one model.

  5. Test it in five minutes.

    1. Paste the block.
    2. Ask a question with two fair answers, like “Should we hire now or wait a quarter?”
    3. Look for a pick, the case against it and what would change its mind.
    4. Read the case against first. Then decide.

The study tested what an agent’s lean does, not this setup. Not a guarantee.