Mar 08,2026 
Most UX researchers who have tried feeding a prompt into an AI tool know the feeling of getting back something useless. A vague summary. A list of generic survey questions that could apply to any product on the planet. The problem is rarely the tool itself. The problem is that the prompt was written the way you might ask a coworker a quick favor, not the way you’d brief a research assistant who knows nothing about your project, your users, or your constraints.
A survey of 100 UX researchers conducted by Lyssna found that 88% identified AI-assisted analysis and synthesis as a top trend for 2026, making it the single most anticipated development in the field. The appetite is there. The skill to use these tools well is still catching up.
What follows are 20 prompts organized by research phase, each written to produce outputs you can actually use, along with the reasoning behind how they’re structured.
Why Structure Matters More Than the Tool You Pick
Nielsen Norman Group published a framework called CARE, which stands for context, ask, rules, and examples. Every prompt should contain those four components to produce a usable output. Context tells the model who you are, what product you’re working on, and what constraints exist. The ask is the specific request. Rules set boundaries on format, tone, length, or scope. Examples show the model what a good output looks like.
This might sound like a lot of work for a single prompt. Nielsen Norman Group acknowledges that directly, noting that with current generative AI tools and models, this amount of strategic context is often necessary to achieve good outputs, and that crafting and iterating on prompts can itself be a time suck. That’s the whole point of templates. You build the structure once, swap in project-specific details each time, and skip the 20 minutes of back-and-forth that produces mediocre results.
Planning Phase Prompts (1 Through 7)
Nielsen Norman Group’s article on planning research with generative AI makes a point worth repeating here. It can be tempting to ask an AI tool to give you a research plan for a project. Don’t do that. Conversational AI can’t give you the assembled final output from a single prompt. Break the plan into components and handle each one separately. They also recommend treating the AI tool as a UX assistant, not a UX mentor, meaning you need to feed it all the steps and details you want it to consider rather than expecting it to ask the right follow-up questions on its own.
Prompt 1: Mapping Assumptions Before Forming Research Questions
Most projects carry assumptions the team has never written down. Those assumptions shape research questions, but if nobody names them first, the questions tend to confirm what everyone already believes.
You are a UX research assistant. I am a [role] working on [product name], which is [brief product description] used by [target user group]. Before we write research questions, I need to surface the assumptions our team is operating under. Here is what we currently believe about our users: [paste 3 to 5 bullet points of team assumptions, e.g., “Users prefer self-service onboarding,” “Most churn happens within the first 14 days,” “Price is the primary barrier to upgrade”]. For each assumption, generate one research question that would test it directly and one that would explore the broader behavior around it. Format the output as a two-column table with the assumption in the first column and the paired questions in the second. Then, at the bottom, add 5 additional research questions about areas our assumptions may be blind to, based on the product description and user group I provided.
This prompt forces the model to work from your team’s specific mental model rather than producing research questions in a vacuum. The two-question pairing, one confirmatory and one exploratory, prevents the common trap of designing a study that only tells you what you already think.
Prompt 2: Selecting Methods Under Real Constraints
Method selection prompts fail when they ignore the practical limits of your team. A recommendation for longitudinal diary studies does not help if your entire study window is two weeks.
Here are the research questions I need answered: [paste 5 to 8 questions]. My constraints are as follows: timeline of [X weeks] from kickoff to final readout, budget of [$X] including participant incentives, team of [X researchers and X note-takers], access to [list tools, e.g., UserTesting, Lookback, Qualtrics], and a participant pool limited to [describe, e.g., “existing users reachable via in-app prompt, no access to churned users”]. Recommend 3 methods. For each one, state which of my research questions it can answer, which it cannot, and what the trade-off is given my constraints. Rank them by the number of questions they address within the timeline I gave. If no single method covers all questions, suggest a combination and explain how to sequence them.
Prompt 3: Writing Screener Questions That Filter Accurately
A weak screener lets the wrong participants into your study. The result is wasted sessions and data you cannot use. This prompt treats screener design as a filtering exercise, not a questionnaire.
I am recruiting participants for a [study type, e.g., moderated usability test] about [topic]. The inclusion criteria are: [paste 3 to 6 specific criteria, e.g., “Has used a project management tool at work in the past 30 days,” “Manages a team of 3 or more,” “Does not work in UX or product development”]. Write a screener with 10 questions. For each question, mark which response qualifies the participant and which disqualifies them. Include at least 2 decoy questions that mask the purpose of the screener so participants cannot guess the “right” answers. Use a mix of single-select and multi-select formats. Do not include any open-ended questions, because this screener will be administered at scale without manual review.
Nielsen Norman Group advises completing inclusion criteria before writing a screener, since the AI can only craft an appropriate filter after it knows who you’re looking to recruit.
Prompt 4: Building a Research Brief for Stakeholder Buy-In
Research plans die in review when the person approving them cannot tell what the study will accomplish in terms they care about. This prompt generates a brief that speaks to product and business outcomes rather than methodology.
Summarize the following research plan in 250 words for a [stakeholder role, e.g., VP of Product] who needs to approve the study. Here is the plan: [paste research questions, methods, timeline, and participant criteria]. The summary should cover what we are testing, which product or business decision it informs, when results will be available, and what risk we carry if we skip this study. Do not include methodology details beyond one sentence. Write in a tone appropriate for a senior leader who reads fast and wants to know what action they need to take.
Prompt 5: Generating Hypotheses Tied to Product Metrics
Research without hypotheses tends to produce descriptive findings that are hard to act on. Tying hypotheses to product metrics ensures the study has a measurable point of comparison.
We are conducting research on [feature or product area] for [product name]. The product metrics we track are: [list 3 to 5 metrics, e.g., “activation rate within first 7 days,” “task completion rate on the dashboard,” “support ticket volume related to billing”]. Based on these research questions: [paste questions], generate 5 testable hypotheses. Each hypothesis should follow this format: “We believe [user segment] will [behavior] because [rationale], and we can measure this through [specific metric or observable behavior].” After listing the hypotheses, flag any research question that does not map cleanly to a product metric and suggest what qualitative evidence would serve as a proxy.
Prompt 6: Estimating Sample Size for Mixed-Method Studies
Researchers regularly underestimate how many participants they need, or they recruit the same number for every study regardless of the method. This prompt asks for the rationale, not a formula.
I am planning a study that uses [method 1, e.g., surveys] and [method 2, e.g., moderated interviews] to answer these research questions: [paste questions]. My target population is [describe, e.g., “enterprise SaaS users in North America, primarily mid-level managers”]. For each method, recommend a sample size with a plain-language explanation for why that number is sufficient. Consider the following: for qualitative methods, explain when I’m likely to reach thematic saturation. For quantitative methods, explain what confidence level and margin of error the sample size supports. If my budget of [$X] limits the sample size, say so and explain what I will lose in coverage.
Prompt 7: Drafting a Recruitment Email That Gets Responses
Recruitment emails written in research jargon get ignored. This prompt prioritizes the participant’s perspective and motivation.
Write a recruitment email for a [study type] about [topic, described in user-friendly terms, not internal product language]. The participant will spend [duration] and receive [$X or gift card] as compensation. The session will be conducted via [platform]. The email is going to [describe audience, e.g., “existing customers who last logged in within 90 days”]. Write the email in a professional but conversational tone. In the first sentence, mention the specific benefit of participating, not the compensation but why their input matters to the product they use. Include a single call to action linking to [placeholder URL]. Keep the email under 150 words. Do not use the phrase “we value your feedback.”
Data Collection Prompts (8 Through 12)
Nielsen Norman Group’s guidance on AI in data collection is measured. AI interviews can work when you need structured feedback at scale. They fall short when topics are complex, specialized, or require flexibility in how questions are asked. The same article advises against using AI tools to moderate usability tests, since the tools are not yet capable of knowing what users are actually doing on screen.
Prompt 8: Writing an Interview Guide That Avoids Leading Questions
Discussion guides often fail because the questions accidentally steer the participant toward the answer the researcher expects. This prompt builds in a bias check.
I am conducting 45-minute moderated interviews with [participant description] about their experience with [feature or workflow]. The research questions driving this study are: [paste 3 to 5 research questions]. Write a discussion guide with the following structure: a 3-minute introduction script that explains the session, sets expectations, and states that there are no right or wrong answers. Then create 5 topic sections, each with a main question and 2 follow-up probes. After the guide, review every question you wrote and flag any that contain leading language, assume a specific behavior, or could be answered with yes or no. Rewrite the flagged questions. End with a 2-minute closing script that invites final thoughts without asking “Is there anything else?”
Prompt 9: Designing Task Scenarios for Unmoderated Tests
Task scenarios written poorly either tell participants exactly what to do, which defeats the purpose, or remain so vague that participants don’t know where to start.
Write 6 task scenarios for an unmoderated usability test of [feature or product]. Participants are [description, e.g., “first-time users who have not completed onboarding”]. Each scenario should describe a realistic goal the user is trying to accomplish without naming specific UI elements, menu labels, or navigation paths. For each task, include a success criterion I can evaluate from the recording and a note about what failure would look like. Avoid instructional phrasing like “click,” “go to,” “select,” or “navigate.” Scenarios should be written in the second person, as if the participant has a real reason to complete the task.
Prompt 10: Building a Survey That Balances Depth and Completion Rate
Long surveys produce drop-offs. Short surveys produce surface-level data. This prompt pushes for an intentional trade-off between the two.
Draft a survey about [topic] for [audience description]. The survey should take no longer than 8 minutes to complete, which means approximately 15 to 18 questions. My research questions are: [paste questions]. For each research question, include at least one survey question that addresses it. Use 5-point Likert scales where opinions are needed, multiple choice where behaviors are being measured, and cap open-ended questions at 2, placing them at the end so they don’t cause early abandonment. Write a brief introduction for the survey that tells the participant how long it will take, what the data is for, and that responses are anonymous. Group questions by theme and place less sensitive questions first.
Prompt 11: Diary Study Prompts That Capture Real Moments
Diary studies fail when the daily prompts are too generic to anchor participants in a specific behavior. If you ask “How was your experience today?” you will get nothing useful.
Create 7 daily diary study prompts for participants who are using [product or feature] over the course of one week. Each prompt should be tied to a specific moment in the participant’s workflow, not a general reflection on the day. For example, do not write “Tell us about your day using [product].” Instead, anchor the prompt in a concrete trigger, such as “The last time you opened [product] today, what were you trying to do?” Keep each prompt under 25 words. Include 1 prompt per day that asks the participant to capture a screenshot or photo of something relevant, and specify what. At the end, add a wrap-up prompt for day 7 that asks the participant to compare their first day to their last.
Prompt 12: Creating Card Labels That Reflect User Mental Models
Card sort results are only as good as the labels on the cards. Labels pulled from internal feature names or product taxonomy will bias the outcome.
Generate 30 card labels for an open card sort about [information architecture topic, e.g., “organizing help center content for a B2B analytics platform”]. Each label should describe a task, action, or piece of content from the user’s perspective, not the product’s internal structure. Avoid product jargon, feature names, and abbreviations. Write each label in 2 to 5 words using language a non-technical user would recognize. After generating the labels, review them and flag any that are ambiguous or could be interpreted in more than one way, then rewrite the flagged labels to reduce ambiguity.
Analysis Phase Prompts (13 Through 17)
AI tools designed for qualitative research analysis can reduce the time spent on transcription, initial coding, and note-taking by 70 to 90 percent, according to figures cited by Lumivero and Looppanel in their respective benchmarks. Those figures refer to the mechanical portions of analysis, not the interpretation. Nielsen Norman Group warns that when you outsource your analysis entirely to AI, you risk more than bad findings. You risk your credibility. A trained researcher considers how a participant’s statement contrasts with other things they said, how the data was collected, and whether the interviewer accidentally primed the participant. AI tools cannot account for those conditions yet.
The practical approach is to use AI for the tedious initial steps while you retain control over interpretation.
Prompt 13: Cleaning a Transcript Without Losing Meaning
Transcripts from automated tools arrive with filler words, false starts, and misattributed speaker labels. Cleaning them manually takes hours per session.
Here is a raw interview transcript: [paste transcript]. Clean up filler words such as “um,” “uh,” “like,” and “you know,” and remove false starts where the participant restarts a sentence. Preserve the participant’s vocabulary and phrasing. Do not paraphrase or condense their statements. If the participant uses informal grammar or unconventional word choices, keep those intact because they carry meaning. After cleaning, write a 200-word summary that captures the 3 to 4 main points the participant made, noting areas where they expressed frustration, confusion, or strong preference. Include 2 to 3 direct quotes that are worth pulling into a highlight reel.
Prompt 14: Generating Initial Codes From Interview Data
Manual coding is the bottleneck in qualitative analysis. Having AI suggest initial codes gives you a draft to react to rather than a blank spreadsheet.
Review this interview summary and its supporting transcript excerpts: [paste summary and key excerpts from one session]. Generate 12 to 18 initial codes based on behaviors described, problems mentioned, workarounds used, and emotional responses expressed. For each code, provide a short name of 2 to 4 words, a one-sentence definition that a second coder could use to apply the code consistently, and one example quote from the transcript. After generating the codes, group any that overlap and suggest merging them. Flag codes that appeared only once and recommend whether to keep or discard them.
Prompt 15: Synthesizing Codes Across Multiple Sessions
Individual session codes are not findings. Findings come from patterns across sessions.
Here are the code sets from [X] interview sessions: [paste all codes with their session labels]. Identify patterns by grouping these codes into 4 to 7 thematic clusters. For each cluster, provide a descriptive name, list the codes that belong to it, indicate how many sessions the cluster appeared in, and write a 2-sentence summary of what the pattern means. If any codes do not fit into a cluster, list them separately and suggest whether they represent edge cases or emerging themes that need more data. Do not force every code into a group.
Prompt 16: Interpreting Split Survey Results
Survey data where responses cluster at extremes or split evenly is harder to interpret than data with a clear pattern. This prompt forces the model to explain the split, not smooth it over.
Here are the results from a survey with [X] respondents about [topic]: [paste summary statistics, including distributions for key questions and any cross-tabulations]. Identify the 3 strongest patterns in the data, meaning the questions where responses clustered most consistently. Then identify 2 areas where responses were split or contradictory. For each split, suggest 2 possible explanations and a follow-up question that could distinguish between them. Do not speculate beyond what the data supports. If the sample size for a particular segment is too small to draw conclusions, say so.
Prompt 17: Rating Usability Issues by Severity With Justification
Severity ratings applied without criteria become arbitrary. This prompt ties each rating to defined impact.
Here is a list of usability issues observed during testing of [product or feature]: [paste issues with brief descriptions of what happened during the session]. Rate each issue on a 4-point severity scale: cosmetic means the problem is noticeable but does not affect task completion, minor means the user is slowed down but completes the task, major means the user fails the task or requires significant effort to recover, and catastrophic means the user is blocked entirely or the error could cause data loss or harm. For each rating, write one sentence explaining why you placed it at that level. If an issue could fall between two severity levels, explain the tiebreaker. After rating all issues, sort them by severity from highest to lowest.
A practical tip when working with any of these analysis prompts: if you’re consistently using over 60% of the context window, break the task down further. Smaller inputs produce more reliable outputs.
Synthesis and Reporting Prompts (18 Through 20)
Prompt 18: Building a Findings Framework With Design Implications
Raw thematic clusters are not a deliverable. Product teams need findings that connect to what they should do next.
Organize these thematic clusters and their supporting evidence: [paste clusters, code counts, and key quotes from each]. Create a findings framework with 3 to 5 top-level findings. For each finding, include a summary statement of one to two sentences written in plain language a non-researcher can understand, 2 to 3 supporting data points drawn from the evidence I provided, a design implication that states what this finding means for the product, and a confidence level of high, medium, or low based on how many data sources support the finding and whether any evidence contradicts it. Do not generate recommendations. I will do that separately with the product team.
Prompt 19: Writing an Executive Summary That Answers the Right Question
Executive summaries fail when they summarize the study instead of answering the question the study was designed to answer.
Write a 250-word executive summary of this research study for [audience role, e.g., “a leadership team reviewing Q3 priorities”]. The study investigated [research question or objective], involved [X participants using method], and ran from [start date] to [end date]. Structure the summary as follows: the first sentence states the single most important finding. The next 2 to 3 sentences provide the supporting evidence. The following sentence explains what this means for the product or business. The final sentence states what the recommended next step is. Do not include a description of the methodology beyond one phrase. Do not open with “This study aimed to…” or any sentence that describes the study rather than its conclusions.
Prompt 20: Prioritizing Recommendations With Rationale
Recommendation lists without prioritization create the illusion that everything matters equally. Product teams need to know what to do first and why.
Here are the design recommendations from this study: [paste recommendations]. Prioritize each recommendation into one of three categories. Quick wins are high impact and low effort, meaning they address a frequent or severe issue and can be implemented within a sprint. Strategic initiatives are high impact and high effort, meaning they require design exploration, engineering investment, or cross-team coordination. Low priority items are low impact or low frequency, meaning few users are affected or the issue is cosmetic. For each placement, write one sentence explaining what makes it high or low impact and what makes it high or low effort. If you lack enough information to categorize a recommendation, say so and state what additional information you would need.
Prompt Chaining: Connecting Each Phase
You’ll notice that many of these prompts reference outputs from earlier steps. That is intentional. Prompt chaining, where the output of one prompt feeds into the next, produces better results than isolated prompts because the model retains context about your project across the sequence. Anthropic’s documentation recommends this approach for complex tasks, noting that breaking a task into focused subtasks improves accuracy and consistency for each step and makes sure the model can fully focus on one subtask at a time.
For a full research cycle, the chain looks like this: Prompt 1 feeds into Prompt 2, which feeds into Prompts 3 and 7, which feed into Prompt 8 or 9, which feed into Prompts 13 through 17, which feed into 18 through 20. Each stage has a specific input, specific instructions, and a specific output.
A Note on Ethics and Data Privacy
Feeding participant data into any AI tool introduces real risks. Unless your organization has a vetted, private deployment, treat public AI services as external by default. Redact identifying information before pasting transcripts.
The regulatory environment around AI and data is tightening. Under the EU AI Act, the penalty framework became operational in August 2025, with fines reaching up to 35 million euros or 7 percent of global annual turnover for violations of prohibited AI practices. The full set of high-risk AI system obligations takes effect in August 2026, according to the implementation timeline published by the European Commission. Separately, South Korea’s AI Framework Act came into effect on January 22, 2026, as reported by the Library of Congress. That law introduces obligations for what it defines as high-impact AI systems in sectors like healthcare, energy, and public services, with the government providing a one-year grace period before imposing administrative fines.
The practical rule: anonymize first, prompt second, verify always.
The Researcher Stays in the Loop
AI tools reduce the time spent on mechanical tasks. The interpretation, the contextual awareness, the judgment call about what matters and what does not, stays with you. These 20 prompts are designed to handle the repetitive, time-consuming portions of each research phase so you can spend your time on the part that actually requires a trained researcher.