> ## Documentation Index > Fetch the complete documentation index at: https://runrehearsals.com/docs/llms.txt > Use this file to discover all available pages before exploring further. # Asking Good Questions > How to write neutral questions, why framing matters even more for AI, and how to ask the same thing more than once ## Why This Page Exists A replica is only as good as the question you ask it. As in a real moderated interview, the quality of your results depends on how the questions are written. There is one added wrinkle for AI. Framing matters more for AI than for humans. Language models can shift their answer on cosmetic rewording even when the meaning has not changed. A leading question does more than bias a replica. It can quietly manufacture the result you were hoping to see. ## How Replicas Fail Differently From Humans Understanding the failure modes makes the anti-patterns below make sense. A leading or loaded question invites agreement. Replicas, like people, drift toward the framing you hand them. On divisive topics, weak questions flatten the real spread. Use larger cohorts and neutral framing to preserve genuine variance. Reveal a price or a friction point in the question and it contaminates every downstream answer. If the question presumes a behavior, replicas will rationalize it rather than tell you it is wrong. ## Anti-Patterns to Avoid | Instead of… | Why it's a problem | Try… | | :-------------------------------------- | :------------------------------------- | :------------------------------------------------------------------------------ | | "How does this make you feel?" | Vague; invites empty sentiment | "What is the first thing you would do after seeing this?" | | "Would you buy this?" | Leading, yes-biased | "You were considering a purchase like this. Walk me through what happens next." | | "Do you love the clean, modern design?" | Loaded and presupposing | "Describe the design in your own words." | | "Is it easy to use and affordable?" | Double-barreled (two questions in one) | Ask about ease and price separately | | "At \$9.99/mo, is this worth it?" | Anchoring on a number | Keep prices out of the question; let willingness-to-pay emerge | | "On a scale of 1 to 10…" | Manufactured precision, no reasoning | Ask for the decision and the reasoning behind it | | "Most people love this. Do you?" | Social-proof anchor | Remove the social proof entirely | ## Rules of Thumb Never state a price you want to test, and never name the weakness you suspect. Let the replica reveal it. Get the honest first reaction before you show options side-by-side. Comparison changes how people think. AI pays extra attention to whatever comes first, so do not list your favored option at the top. Offer balanced go and no-go choices, not a menu tilted toward the answer you want. ## Asking the Same Thing More Than Once Yes, you can, and there are two deliberate ways to do it. ### 1. The rephrase (fragility) check Rewrite one or two of your most important questions in slightly different words and run them. If a replica gives a noticeably different answer to a reworded version that means the same thing, the original question was fragile: the framing, not the substance, was driving the result. Robust findings survive rewording. A good gut check: if you rephrased this question to bias the other direction, would you notice the difference? If yes, rewrite it toward neutral. ### 2. The split by intent Run the same concept against the same cohort with two different mindset framings (for example, two different jobs-to-be-done). Comparing the results shows how the concept lands for different motivations, which is useful signal rather than a contradiction. ## Screening First If a study is only meaningful for a subset of people (for example, people who personally pay for the subscription), screen for it. Screening keeps you from surveying a fantasy population, which is the most common way a result ends up looking impressive but wrong. You add up to three pre-screen questions when you [build a cohort](/docs/platform/building-cohorts); each replica is then marked Qualified or Rejected, with the confidence, the reasoning, and a supporting quote shown. ## We Take a First Pass, but Review Before You Run When you set up a Rehearsal in the [intake chat](/docs/platform/new-rehearsal), Rehearsals takes a first pass at your questions and runs them through our system to reduce leading, loaded, or anchored framing. That first pass is a starting point, not a guarantee, so always read the questions yourself before you launch. What to do when a simulation does not match your expectations.