Question Design: Do's & Don'ts

Do’s

A well-defined scale or option set gives AI twins a clear decision space.

Likert Scales

Always define the full range and label both endpoints.

ExamplesHow satisfied are you? 1 = Very dissatisfied … 5 = Very satisfied.
How likely to recommend? 1 = Not at all … 10 = Extremely likely.

Multiple Choice / Rank Order

Write out every option, make them mutually exclusive and at the same level of abstraction. Offer “Other” when useful. Avoid numbering or lettering options (1, 2, 3 or A, B, C) — use descriptive labels.

ExamplesWhich products do you use? Select all that apply: Product A, B, C, other.
Rank your top 3: quality, price, design, brand, availability.

Tell the twin what kind of answer you expect.

Numeric Answers

State the unit and the format.

ExamplesHow much would you pay? A single number in CHF.
How many times per week? Answer with a whole number.

Free-Text Answers

State the length and angle.

ExamplesIn 2–3 sentences, describe what you like most and why.
In one short paragraph, explain the main reason behind your choice.

Don’ts

Double-Barreled Questions

Two assessments in one question.

Avoid

How satisfied are you with the price and quality?

Better

Split price and quality into two separate questions.

Mixed-Level Answer Options

All options should be at the same level of abstraction.

Avoid

Which feature do you like? Feature A / Feature B / I don’t like this product

Better

Separate the feature preference question from a general attitude question.

Loaded Questions

Embedded value judgments bias answers. LLMs are particularly susceptible.

Avoid

How much did you enjoy our award-winning service?

Better

How would you rate your most recent service experience?