A survey item must state the behavior or opinion it intends to measure, the reference period, and terms respondents can interpret consistently. Ask the model to compare candidate wording with that construct, flag leading adjectives, double questions, hidden assumptions, and ambiguous time spans. The model's preferred wording remains a draft until reviewed with people who resemble the respondents. If a historical trend requires unchanged wording, label the trade-off rather than quietly editing the item and treating the results as comparable. Keep the final question text and version in the data record.
Make this comfortable
Survey prompts: check wording against the intended construct
Operational case
Meridian initially drafts, 'How frustrating is our confusing pickup window?' That wording presumes both confusion and frustration. The revision asks, 'During your most recent pickup, how clear was the time window shown before confirmation?' with a specified response scale. A separate item asks whether the person could plan around that window; clarity and planning are not forced into one answer. The assistant records the wording change, but it cannot claim that the new item is unbiased without a respondent pilot.
Construct A: perceived clarity at most recent pickup
Draft: 'How frustrating is our confusing pickup window?'
Revised: 'How clear was the time window shown before confirmation?'
Construct B: ability to plan -> separate item
Question version: Q7-v2Performance and review cost
Reviewing Q items against a construct ledger is O(Q) for a first pass, with human testing adding respondent time. A short question can be expensive if it changes what is measured and invalidates a trend. Preserve the exact text and order at fielding, since a paraphrased label in the report cannot substitute for the instrument respondents actually saw.
Common Mistakes
- Do not embed the desired answer in an adjective.
- Do not combine clarity and planning into one response.
- Do not compare two waves after changing question wording without marking the break.
Connected lessons
- Prompt engineering applications
- Prompt Engineering
- Clarification gates: ask only when a missing fact changes the outcome
- Multilingual prompts: test policy meaning across languages
- Survey prompts: define the population and invitation frame
- Survey prompts: validate answer options and skip paths
- Survey prompts: pilot the instrument before fielding
- Survey prompts: reconcile invitations, submissions, and item denominators
- Survey prompts: code open text without exposing respondents
- Survey prompts: release findings with explicit scope
- Project: review Meridian pickup survey findings
- Survey prompt decisions
prompt engineering
survey research