AI Prompts for UX Designers (2026)

Every prompt below has [BRACKETED] placeholders — replace them with your actual artifacts before sending. The single biggest quality lever is grounding: paste your real interview transcript, your real screen flow, or your real copy, rather than describing it. A model synthesizing your actual notes is doing analysis; a model inventing plausible UX findings from a one-line brief is doing fiction.

Ask for structure. When you want themes, request a table; when you want a heuristic review, request a numbered finding-per-issue format with severity. Asking the model to think step by step before answering improves multi-step reasoning, a pattern grounded in Chain-of-Thought prompting (Wei et al., 2022) — see our chain-of-thought prompting guide for when it helps and when it just adds noise.

For deeper work, set a role and constraints in a reusable system prompt rather than re-typing context every time — our how to write a system prompt walk-through covers the pattern.

**1. Theme extraction from interviews** — "You are a UX research analyst. Below are [N] user-interview transcripts about [PRODUCT/FEATURE]. Cluster the findings into 4-7 themes. For each theme give: a one-line name, the count of participants who mentioned it, 2-3 verbatim quotes (copied exactly, not paraphrased), and the design implication. Return as a table: Theme | Participants | Quotes | Implication. Base everything only on the transcripts below — do not invent quotes, counts, or findings. Where evidence is thin, say so. Transcripts: [PASTE]."

**2. Affinity map from raw notes** — "Act as a facilitator running an affinity-mapping session. Group the sticky-note observations below into clusters, name each cluster, and flag any observation that contradicts another. Then list the top 3 'How might we' questions the clusters suggest. Use only the notes provided; mark anything you inferred with [INFERRED]. Notes: [PASTE]."

**3. Survey open-text coding** — "Code the open-ended survey responses below into a tag taxonomy. First propose 6-10 tags with definitions, then apply them, then return a frequency table sorted high to low. Keep tags mutually exclusive where possible and note any response that fits none. Do not fabricate sentiment scores or percentages beyond the actual counts. Responses: [PASTE]."

**4. Microcopy variants for a UI element** — "Write 5 variants of [ELEMENT — e.g. an empty-state message / error message / button label] for [PRODUCT]. Context: the user just [WHAT HAPPENED] and we want them to [DESIRED ACTION]. Voice: [3 words]. Constraints: under [N] characters, plain language at a grade-7 reading level, no blame, no jargon. For each variant give the text and a one-line note on the tone it strikes. Lead with what the user can do next."

**5. Error messages that help** — "Rewrite these error messages so each one says what went wrong in plain terms, why it matters to the user, and the single next step. Remove system codes and blame language ('you entered an invalid…'). Keep each under [N] characters. Return Original | Rewrite | What changed. Messages: [PASTE]."

**6. Onboarding tooltip sequence** — "Draft a [N]-step onboarding tooltip sequence for [FEATURE]. Each tooltip: a short header (max [N] chars), one sentence of body, and a verb-first CTA. Progressive disclosure — reveal only what's needed at that step. Avoid 'click here' and 'simply'. Map each tooltip to the screen element it points at. Feature flow: [DESCRIBE OR PASTE]."

**7. Nielsen heuristic evaluation** — "Act as a usability expert. Evaluate the flow described below against Nielsen's 10 usability heuristics. For each heuristic that is violated, give: the heuristic name, the specific issue, a severity rating (0-4, where 4 = usability catastrophe), and a concrete fix. Skip heuristics with no issue. Be specific to this flow, not generic. Note where you'd need to see the live product to judge. Flow: [DESCRIBE SCREENS, STATES, AND COPY — or paste the spec]."

**8. Cognitive walkthrough** — "Run a cognitive walkthrough for this task: [TASK GOAL]. For each step, answer the four standard questions: (1) Will the user try to achieve the right effect? (2) Will they notice the correct action is available? (3) Will they associate the action with the effect? (4) Will they see progress after acting? Flag every step where the answer is 'no' or 'maybe' and suggest a fix. Steps: [LIST THE INTENDED STEPS]."

**9. Accessibility first-pass review** — "Review the screen described below for common accessibility issues against WCAG 2.2 AA at a heuristic level: color-contrast risks, missing labels, touch-target size, focus order, and reliance on color alone to convey meaning. Return Issue | WCAG criterion | Why it matters | Suggested fix. Note clearly that this is a heuristic pre-check, not a substitute for testing with assistive technology and real users. Screen: [DESCRIBE]."

**10. Evidence-based persona** — "Build a single UX persona from the research below. Include: goals, top 3 frustrations, context of use, the job-to-be-done in the user's own words, and one anti-goal (something they explicitly don't want). Ground every attribute in the supplied evidence and cite which note supports it; mark anything you inferred with [INFERRED]. Do not invent demographics, quotes, or statistics. Research: [PASTE]."

**Bonus — card-sort to IA** — "Here are the results of an open card sort: [PASTE GROUPINGS]. Propose an information architecture: a top-level navigation of 4-7 items with child pages, and label each grouping with a name users would recognize (not internal jargon). Flag items that participants placed inconsistently and explain the trade-off. Then suggest 3 tree-test tasks to validate the structure."

Turn the persona output into a reusable profile with the Customer Persona Generator, and if you're presenting findings, draft the deck with the Presentation Outline Generator.

Never paste raw participant data containing real names, emails, recordings, or any other personally identifiable information into a public chatbot — anonymize transcripts first. Treat the model as a synthesis assistant, not a source of truth: it will happily invent a quote, a participant count, or a 'study shows' statistic if your prompt leaves room for it, so always add 'use only the supplied material; do not fabricate' and verify counts yourself.

Don't outsource judgment. AI can flag a heuristic violation, but severity and prioritization depend on your users and business context. And remember the model can't see your live product — a flow described in text is a proxy; confirm real issues with moderated usability testing. For the security side of pasting untrusted content into prompts, see the prompt injection defense checklist.

For long research transcripts and synthesis, a long-context frontier model handles the volume best — Claude Opus 4.8, GPT-5.5, or Gemini 3.5 Pro all do well; see our how to choose an AI model guide for the trade-offs. For high-volume microcopy variants where cost matters, an efficiency-tier model like Claude Haiku 4.5 or Gemini 3.5 Flash is plenty. The table below compares durable dimensions rather than prices — check live rates at the linked official pages.