User research is heavy on unstructured work: long transcripts, sprawling spreadsheets, half-finished personas, and sticky notes that lost their adhesive three sprints ago. That is exactly the kind of work large language models handle well. With the right prompts, an LLM becomes a research assistant that drafts interview scripts, summarizes transcripts, surfaces themes, and writes the report you've been avoiding — without replacing the judgment that makes research valuable in the first place.

This guide collects reusable AI prompts for the core stages of UX research: planning, interviewing, survey design, synthesis, usability testing, reporting, and persona work. Each prompt is structured so you can swap in your own context and get usable output on the first try. Use them with ChatGPT, Claude, Gemini, or any capable model. Pair them with Promptwright to keep your prompts versioned, shared across the team, and ready to run.

Why AI fits UX research

UX research produces a lot of natural language. Transcripts, field notes, open-ended survey responses, support tickets, app store reviews — all text. LLMs are trained on text, so they are genuinely useful here in ways they are not for, say, rendering a Figma file. Specifically, models can:

Summarize long transcripts into structured notes without losing nuance.
Cluster qualitative responses into themes, faster than manual affinity mapping.
Draft first versions of interview scripts, surveys, and screeners that you refine rather than start from scratch.
Translate research outputs into stakeholder-friendly formats: one-pagers, slide outlines, executive summaries.
Critique your own work — pointing out leading questions, biased wording, missing segments.

What models cannot do is decide whether your research question is worth asking, whether your sample is representative, or whether a finding is strategically meaningful. You still own those calls. The prompts below are scaffolding, not autopilot.

A note on privacy before you paste anything in

Before using any of these prompts with real participant data, think about what you are sending to a third-party model. Raw transcripts often contain names, employer names, health details, or other identifiers. Good practice:

Redact participant names and contact details before pasting. Replace with P1, P2, etc.
Check your organization's data policy and the model provider's terms. Many teams route sensitive work through enterprise plans with no-training guarantees, or self-hosted models.
Aggregate where possible — summarize batches of responses rather than dumping verbatim text.
Ask consent in your research agreement for transcript processing by AI tools. This is increasingly standard and only takes a sentence.

1. Planning the study with AI

A sloppy research plan is the most expensive mistake in UX research, because every downstream artifact inherits its flaws. Use AI to pressure-test your plan before you recruit a single participant.

Prompt: stress-test research questions

You are a senior UX researcher reviewing a study plan. Be critical, specific, and concise.

Context:
- Product: [one-line description, e.g. "B2B analytics dashboard for marketing ops teams"]
- Research goal: [what decision this study should inform, e.g. "whether to rebuild the report builder or improve it"]
- Draft research questions:
  1. [question]
  2. [question]
  3. [question]
- Participants: [who, how many, how recruited]

Tasks:
1. For each research question, say whether it is leading, answerable, and likely to produce actionable insight. Score each 1-5.
2. Identify any missing question that should be added.
3. Flag any question likely to produce vague or social-desirability-biased answers.
4. Rewrite any question that scores below 4.
5. Suggest 2-3 probes or sub-questions for each surviving question.

Output as a table followed by rewritten questions.

Prompt: write a screening questionnaire

Write a screener questionnaire for a UX research study.

Study details:
- Product: [description]
- Research goal: [goal]
- Target participant profile: [role, seniority, behaviors, e.g. "marketing operations manager who has rebuilt at least one dashboard in the last 6 months"]
- Disqualifiers: [e.g. "works at a competitor", "has never used a dashboarding tool"]
- Study format: [60-min remote interview, async diary study, etc.]
- Incentive: [$ amount]

Produce 8-12 screening questions. For each:
- The question text, in plain language a recruiter can read aloud
- Why you are asking it (one sentence)
- The response options
- The qualifying answers vs. the disqualifying answers
- Any trap question to detect fake or professional respondents

Avoid asking directly for the qualifying trait (e.g. don't ask "Do you use dashboards?"). Use behavioral proxies instead.

The "avoid asking directly" instruction is the secret. Professional survey-takers learn to spot qualifying answers. AI will reliably produce traps if you ask it to.

Prompt: draft a research brief

Draft a one-page research brief for stakeholders.

Inputs:
- Decision this study supports: [decision]
- Research questions (already validated): [list]
- Method: [method, duration, sample size]
- Participants: [profile, n=number]
- Outputs: [interview recordings, transcript summaries, persona updates, etc.]
- Timeline: [start date, end date, readout date]

The brief should answer, in order: What decision are we informing? What will we ask? Who will we talk to? What will we deliver? When? Include a "what this study will NOT answer" section to manage scope.

Keep under 400 words. Plain business language, not research jargon.

2. Interview scripts that don't lead

A leading question is worse than no question, because it feels productive. AI is genuinely good at detecting leading wording because it has seen millions of survey items. Use it as a backstop.

Prompt: generate an interview script

Write a semi-structured interview script for a 60-minute remote session.

Study context:
- Product: [description]
- Research goal: [goal]
- Research questions (validated):
  1. [RQ1]
  2. [RQ2]
  3. [RQ3]
- Participant profile: [role, seniority]
- Likely prior knowledge of product: [none / light / power user]

Structure:
1. Intro and consent (2-3 min) — write the actual script the researcher reads.
2. Warm-up questions about their work context (5-7 min).
3. Core questions mapped to each RQ — list 2-4 questions per RQ. Mark the primary one.
4. Probes for each core question — short follow-ups to use if the participant doesn't elaborate.
5. A demo or retrospective task if appropriate — describe it.
6. Closing question: "what should we have asked that we didn't?"

Constraints:
- No leading questions.
- No binary yes/no questions; prefer "tell me about a time" or "walk me through" framing.
- Order questions from broad context to specific, with the most important RQ covered first.
- Mark estimated minutes for each section.

Prompt: rewrite leading questions

You are an expert qualitative interviewer. I will paste interview questions. For each one:

1. State whether it is leading, double-barreled, or assuming.
2. Explain the problem in one sentence.
3. Provide two rewritten versions: one neutral open version and one story-eliciting version ("tell me about a time when…").

Paste as a list. Be ruthless.

Questions:
[questions here]

Run this on every script before you go to field. It catches wording you've gone blind to.

3. Survey design that produces usable data

Prompt: write a mixed-methods survey

Design a survey for [audience]. Research goal: [goal]. Length: no more than [N] questions, completable in under [X] minutes.

For each question, output:
- Q[number]: [question text]
- Type: [multiple choice / Likert / open text / matrix / ranking]
- Response options (if applicable)
- The research question this answers
- A note on analysis: how this response will be analyzed

Rules:
- Avoid leading wording and loaded adjectives.
- Use 5-point Likert scales with balanced anchors (Strongly disagree → Strongly agree) and include a neutral midpoint.
- Open-text questions only where a quantitative item cannot capture the answer — max 3.
- For multiple-choice, list exhaustively and include "Other (please specify)" where appropriate.
- Randomize order suggestion for any matrix block where order effects are likely.
- End with one open question: "Anything else we should know?"

Also produce a short "analysis plan" section explaining how each question maps to the research goal and what chart or statistic will be used to summarize it.

Prompt: review a draft survey

Critically review this draft survey for [audience, goal]:

[paste survey]

For each item, check:
- Leading or biased wording
- Ambiguous language
- Mutually exclusive and exhaustive response options
- Double-barreled questions
- Order effects across items
- Missing "don't know" or "not applicable" where needed
- Any item where social desirability will skew responses

Output a table: Question | Issue | Severity | Fix. Then suggest a revised order for all items.

This works because models have ingested thousands of survey methodology guides and bad real-world surveys. The checklists come out reliably.

4. Synthesizing transcripts and notes

Synthesis is where AI pays for itself. A 60-minute transcript is 8,000-12,000 words — three to five of them per study is more than anyone can hold in their head. Models are good at this if you constrain them.

Prompt: summarize one transcript

Summarize this interview transcript into a structured research note.

Transcript (participant ID: P1, role: [role]):

[paste redacted transcript]

Output:
1. Participant snapshot — 2-3 sentences on their context and how they relate to the product.
2. Key quotes — 4-6 verbatim quotes that illustrate their main points. Each must be a direct quote from the transcript, in quotation marks, with a one-line interpretation underneath.
3. Findings — 3-6 bulleted findings, each tagged with the research question it speaks to (RQ1, RQ2, etc.).
4. Surprises — anything the participant said that contradicted hypotheses or other participants, or that warrants follow-up.
5. Open questions for the researcher — what to probe in later interviews.

Do not invent quotes. If a quote is partial or paraphrased, mark it [paraphrased].

The "do not invent quotes" constraint matters. Models will happily fabricate plausible-looking quotes that fit the theme. Requiring verbatim extraction with the [paraphrased] tag keeps you honest. Spot-check every quote against the transcript.

Prompt: thematic synthesis across participants

You are doing thematic synthesis across [N] interview summaries. I will paste them.

Method:
1. Identify recurring themes that appear across at least 2 participants.
2. For each theme, give it a short name, a one-sentence definition, and list which participants contributed to it (by ID).
3. Note any participant who actively disagreed with or complicated the theme.
4. Rank themes by frequency and by salience (how strongly participants emphasized them).
5. Produce a "disconfirming evidence" section: things participants said that push against the dominant themes.

Output a theme table, then 2-3 paragraphs interpreting the patterns. Then produce 5-7 design implications as "If [theme] is real, we should…" statements.

Paste summaries below:
[paste]

The "disconfirming evidence" block is the most important one. Models naturally converge on the majority pattern. Forcing them to surface what doesn't fit prevents you from cherry-picking supportive quotes — the cardinal sin of qualitative research.

Prompt: affinity-diagram draft

Take these [N] participant quotes and findings and produce an affinity map text outline.

For each top-level cluster, produce:
- Cluster name (short, evocative)
- Sub-themes within the cluster (2-5)
- The specific quotes or findings that fall into each sub-theme, with participant IDs

Aim for 4-7 top-level clusters. Group by meaning, not by question asked.

Paste data:
[data]

You can paste this straight into FigJam or Miro afterwards.

5. Usability testing

Prompt: write a usability test script

Write a moderated usability test script for [product], testing [task or specific flow].

Participant profile: [role, familiarity level]
Session length: [minutes]
Modality: [remote screen-sharing]

For each task, produce:
- Task scenario (no leading instructions — describe the situation, not the click path)
- The success criteria (what counts as completion)
- The critical steps to watch for
- Likely failure points and what to probe if the participant gets stuck
- A post-task question (single ease question: "On a scale of 1-7, how difficult was this task?")

Open with a brief pre-test orientation script (set expectations, get consent for recording, remind them you're testing the product not them).
Close with a post-test debrief: 3 open questions about overall impression, most frustrating moment, and what they'd change.

Prompt: analyze a usability test session

Summarize this usability test session.

Participant: P[1], [role], [familiarity level].
Tasks tested: [list]

Transcript / notes / observer log:
[paste]

Output:
1. Per-task table: Task | Completed? (yes/no/partial) | Time | Errors | Severity (blocker / major / minor) | Verbatim quote
2. Top 3 usability issues ranked by impact, each with:
   - What happened
   - Why it happened (root cause hypothesis)
   - Affected participant count if known (e.g. "P1, P3, P4")
   - Recommended fix priority (P0 / P1 / P2)
3. What worked well (so it doesn't get lost).
4. Quotes worth putting in the readout deck.

6. Personas and journey maps

AI can draft personas from real research data — but only from real research data. Asking it to "make up a persona for marketing managers" produces fiction. Use it on synthesized findings.

Prompt: persona draft from research

Based on these [N] participant summaries, draft 2-3 personas that represent the distinct segments I observed. Do not invent traits — derive everything from the data.

Input summaries:
[paste]

For each persona:
- Name and role (use a representative label, not a fictional name with a photo)
- Goals (2-4, grounded in quotes)
- Frustrations and pain points (2-4)
- Behaviors and workarounds
- A "day in the life" paragraph, grounded in observed behavior
- Representative quotes from the data, with participant IDs
- Which other personas this one differs from, and how

End with a "what this persona is NOT" note — what assumptions should we not make about this user?

Prompt: journey map outline

Draft a journey map for [persona] going through [scenario, e.g. "evaluating and signing up for our tool"].

Stages to cover: [awareness / evaluation / signup / onboarding / first value / ongoing use / churn-risk] — adjust to fit.

For each stage:
- What the user is trying to accomplish
- Touchpoints used
- Their emotional state (with intensity 1-5)
- Pain points observed in research
- Moments of delight
- Opportunities for the product

If the data is thin for a stage, say "Insufficient evidence — needs follow-up research" rather than filling it in.

Findings to ground this in:
[paste]

7. Reporting and stakeholder comms

Good research that nobody reads is wasted. Use AI to ship reports that actually get consumed.

Prompt: stakeholder one-pager

Write a one-page stakeholder summary of this research study. Audience: [executives / product leadership / engineering leads]. Decision at stake: [decision].

Inputs:
- Study scope: [n=X, method, dates]
- Top findings (3-5): [list, with rough participant support]
- What surprised us: [1-2 things]
- Recommended next steps: [list]

Format:
1. Headline finding (one sentence, quotable).
2. TL;DR — 3 bullets max.
3. Findings — each with a 1-sentence claim + 1 short supporting quote.
4. So what? — what should we do now, concretely.
5. What we still don't know.

Plain English. No research jargon. Under 300 words.

Prompt: slide deck outline

Produce a slide-by-slide outline for a 15-minute research readout. Audience: [product leadership]. Decision at stake: [decision].

Findings:
[paste]

For each slide:
- Slide title (declarative, not topical — e.g. "Onboarding takes 3x longer than we assume", not "Onboarding findings")
- 1-2 bullets of supporting evidence
- Suggested chart or visual
- Speaker note (what to say, not what's on the slide)

Aim for 8-12 slides. Open with the "so what" headline, not methodology. Methodology goes on slide 2.

Patterns that make these prompts work

A few things show up across all of them. Internalize these and your own prompts will improve:

Always give context first. Product, goal, audience, constraints. Models default to generic. Specificity breaks the default.
Specify output format. Tables, bullets, sections, word counts. "Summarize" without a structure produces mush.
Map output to the research goal. Every prompt above ties findings back to research questions or decisions. This stops the model from drifting into interesting-but-irrelevant observations.
Force disconfirming evidence. Models trend toward agreement. Asking them to surface contradictions, edge cases, and what's missing is the cheapest bias control you have.
Forbid invention. Explicit "do not invent quotes / data / participants" instructions cut hallucinations dramatically. Add [paraphrased] tags as an escape valve so the model doesn't silently paraphrase while pretending it's verbatim.
Decompose synthesis. Don't ask one giant prompt to summarize ten transcripts at once. Summarize each transcript first, then run synthesis on the summaries. Two-stage prompts are more accurate and let you audit intermediate output.

Mistakes to avoid

Pasting raw transcripts with PII. Redact first. Always.
Trusting AI-surfaced themes without re-checking against the data. Use AI to draft, then verify the quotes are real and the pattern actually holds across participants — not just the three it cherry-picked.
Letting AI write your recommendations without owning them. Recommendations are a strategic call, not a summary. Use AI to draft options, then you decide.
Skipping the disconfirming-evidence prompt. This is the single highest-value addition to any synthesis prompt. Do not skip it.
Treating AI persona output as personas. If it isn't grounded in real research data, it's a marketing fiction. Use the prompt that grounds in findings, not one that asks the model to invent a user.

Putting it together

A realistic workflow for a small study might be:

Run the plan stress-test prompt on your draft research questions.
Run interview script prompt; then run rewrite leading questions on the result.
Run screener prompt; give it to your recruiter.
Conduct interviews with the script in front of you (you don't need AI during the session).
After each interview, run summarize one transcript on the redacted transcript.
After all interviews, run thematic synthesis across participants on the summaries — including the disconfirming-evidence block.
Run persona draft from research and journey map outline from the same findings.
Run stakeholder one-pager and slide deck outline outputs.
Sanity-check every quote against the source transcript before you ship.

Used this way, AI does what a junior research assistant would have done in a parallel universe — drafts, summarization, first-pass synthesis. You keep the judgment work: study design, sampling, interpretation, recommendations. The result is more studies shipped, faster readouts, and fewer late nights turning transcripts into something stakeholders will actually read.

Want a versioned library of these prompts you can edit, share with your team, and trigger from a single workspace? Sign up at Promptwright and we'll get you started.

AI Prompts for UX Research: Templates for Interviews, Surveys, and Synthesis