write-survey

Run every question through a bias lint covering acquiescence, social desirability, framing, double-barreled items, prediction, unbalanced scales, non-exhaustive options, missing opt-outs, order effects, and fatigue. One violation means rewrite, then soft-launch to a small sample before the full send.

Use semantic differential by default when rating a thing along a dimension like easy versus difficult, since it reduces acquiescence and social-desirability bias. Reserve Likert agree-disagree statements for genuine attitude constructs that cannot be phrased as a bipolar pair.

No. Standardized instruments like SUS, SEQ, NPS, CSAT, and CES must ship verbatim because their only value is score comparability against benchmarks. Reworded SUS is not SUS, so choose verbatim or fully custom, never in between.

People cannot predict their own behavior, so stated-intent answers are noise. Ask about past or current behavior instead, such as whether they paid for something similar in the last three months, or flag any surviving intent item as under-evidenced rather than reporting it as demand.

A survey measures attitudes and self-reported behavior at scale; it cannot observe what people actually do or explain why. If the research question is about why something happens or whether users can complete a task, use interviews or usability testing instead.