How to Use AI Prompts for UX Research and User Interview Synthesis

Twelve user interviews. Forty hours of transcripts. Your PM wants themes by Monday. You paste everything into Claude and get polished themes that sound right — because the model invented consensus where participants disagreed. UX research synthesis with AI works as a prompt chain aligned to qualitative method — not as "summarize all interviews." Braun & Clarke's six phases map cleanly to sequential prompts with human review gates. This guide covers the 6-phase prompt chain, code → theme → quote extraction, frequency notation (8 of 15), automation bias warnings, affinity mapping prompts, JTBD connection, human review requirements, and FAQ. Human review is required at every phase gate. AI accelerates coding; researchers own interpretation.

Automation bias warning

Polished AI themes feel true because they're fluent. Researchers over-trust synthesized output and skip re-reading raw transcripts — the main failure mode in 2026 UX AI workflows. Mitigations: - Frequency notation on every theme (8 of 15 participants, not "many users") - Verbatim quote requirement with participant ID - Researcher must approve themes before JTBD or roadmap connection - Ban: "users want" without frequency + quote If you only read AI output, you did not synthesize — you summarized a summary.

Braun & Clarke six phases → prompt chain

Phase 1 Familiarization → read transcripts, note early impressions (human + optional AI summary per transcript) Phase 2 Initial coding → code line-by-line or segment Phase 3 Theme search → group codes into candidate themes Phase 4 Theme review → check themes against coded data and full transcripts Phase 5 Theme definition → name themes, write definitions, select exemplar quotes Phase 6 Report → write findings with frequencies and quotes One mega-prompt for all six phases → coherent nonsense. Chain prompts with human gates between phases.

Phase 1 — Familiarization prompt

TRANSCRIPT: [single interview, participant P03] TASK: Familiarization notes only — no themes yet. OUTPUT: - 5 bullet observations (behaviors, surprises, emotional moments) - Initial memos: phrases that stand out (verbatim, max 3 quotes, 25 words each) - Do not generalize beyond this participant Researcher reads transcript + AI notes. AI notes are hints, not findings. Claude Opus 4.8 for nuanced emotional tone. Run per transcript, not batch.

Phase 2 — Initial coding prompt

TRANSCRIPT: [P03] CODEBOOK (initial — inductive): open coding, short descriptive labels RULES: - Code segment by segment (speaker turn or paragraph) - Format: [code] | segment excerpt (max 20 words) | line ref - Multiple codes per segment allowed - No theme names yet — codes only OUTPUT: code table Researcher spot-checks 20% of codes against transcript. Adjust codebook before phase 3.

Phase 3 — Theme search prompt

CODED DATA: [paste all codes from all participants with P IDs] TASK: Group codes into candidate themes (semantic similarity) OUTPUT: Candidate theme | constituent codes | participant IDs represented RULES: - Theme must include codes from ≥2 participants OR flag SINGLE-PARTICIPANT theme - Do not merge codes that contradict Frequency preview: count distinct P IDs per candidate theme.

Phase 4 — Theme review prompt

CANDIDATE THEME: [name] CODES: [list] FULL TRANSCRIPTS: [paste or attach relevant excerpts]

TASK: Does this theme hold? Contradicting evidence?

OUTPUT:

• Fit assessment: strong / partial / reject
• Contradicting quotes with P ID
• Revised theme name if needed

Human decision: accept, revise, or reject theme. AI proposes; researcher disposes.

Phase 5 — Theme definition prompt

APPROVED THEMES: [list from human gate]

TASK: For each theme write 2-sentence definition + 2 exemplar quotes

RULES:

• Quotes verbatim under 30 words with P ID
• Frequency: X of N participants (N = total study N)
• Definition describes pattern, not solution

OUTPUT: theme card per theme

Example frequency: 8 of 15 participants described checkout confusion (P02, P07, P09...).

Phase 6 — Report prompt

THEME CARDS: [paste approved]

STUDY CONTEXT: [product, method, N, dates]

AUDIENCE: [PM / design / exec]

OUTPUT:

1. Study summary (method, N, limitations)
2. Findings (theme order by prevalence)
3. Each finding: frequency notation + 2 quotes + implication (not solution)
4. Limitations and what we cannot claim

BAN: roadmap recommendations unless tagged HYPOTHESIS for team discussion

Frequency notation — mandatory format

Always: X of N participants — never "most users" or "many participants."

Prompt rule: Every theme sentence starts with frequency or includes (X/N) immediately after theme name.

Example: Checkout distrust (9/15) — participants hesitated before entering card data...

If X=1, label anecdotal unless intentional single-case study.

Quote extraction rules

Verbatim only — no smoothing grammar.

Max quote length 30 words unless longer needed — then ellipsis with [...] mid-quote.

Format: "quote text" — P07

Prompt: If exact wording uncertain from transcript, mark [PARAPHRASE — VERIFY] and researcher fixes.

Never attribute quote to wrong participant.

Affinity mapping prompt

After phase 3, for virtual affinity:

INPUT: all codes with P IDs as sticky text list

TASK: Cluster into groups. Name each cluster provisionally.

OUTPUT: cluster | codes | P ID count

Researcher rearranges in FigJam/Miro — AI cluster is starting layout, not final map.

Prompt: Allow overlapping codes in multiple clusters if researcher will split later.

JTBD connection prompt (post-synthesis)

APPROVED FINDINGS: [theme cards with frequencies]

TASK: Map themes to JTBD framework — Jobs, Pains, Gains

RULES:

• Each JTBD item links to theme + (X/N) + quote
• Tag speculative jobs as HYPOTHESIS — not validated by study scope
• Do not invent jobs not grounded in themes

OUTPUT: JTBD table | evidence theme | frequency

PM uses for opportunity framing — researcher validates before roadmap.

Multi-interview batch coding

Do not paste 15 transcripts in one coding prompt — context collapse and participant conflation.

Loop: one transcript → codes → append to master code table.

Master table format: code | excerpt | P ID | line ref

Merge in phase 3 only.

Before vs after synthesis

Weak:

Summarize these user interviews and give me themes.

[15 transcripts]

Output: 5 generic themes (ease of use, performance), no frequencies, blended quotes, false consensus.

Strong:

6-phase chain, frequency on every theme, verbatim quotes, human gates, automation bias checks.

Output: 4 validated themes with 8/15, 6/15 counts, dissent noted, limitations section.

Dissent and negative cases

Prompt phase 4:

For each theme, list participants who discussed topic but DO NOT fit theme — negative cases.

Include negative case quotes in theme card if partial fit.

Dissent prevents false unanimity.

Researcher review checklist

• Every theme has X/N with correct N?
• Quotes verified verbatim in transcript?
• P IDs never mixed up?
• Single-participant themes flagged?
• AI recommendations separated from findings?
• Limitations section present?
• Researcher read ≥20% raw transcript spot-check?

Sign-off: researcher name + date before shareout.

Model routing

Claude Opus 4.8: coding nuance, theme review, quote fidelity

GPT-5.5: structured code tables JSON for research repositories

Gemini 3.1 Pro: long single transcripts in familiarization

Gemini 3.5 Flash: initial pass highlight extraction — recode on Opus

Final report always Opus or human-written — not Flash.

JSON codebook export

For Dovetail / research repo integration:

{ "codes": [{ "id", "label", "excerpt", "participant", "line" }], "themes": [{ "name", "frequency", "participant_ids", "quotes" }] }

Same frequency and verbatim rules.

Stakeholder-specific report overlays

| Audience | Emphasis |

| PM | JTBD mapping, frequency, no design pixels |

| Design | Behavioral quotes, task failure moments |

| Exec | Top 3 themes by prevalence, business risk framing |

Same theme cards — different report prompt overlay.

Common failures

Single mega-prompt synthesis → false themes

No frequency notation → overstated prevalence

Paraphrased quotes → misattributed pain

Skipping phase 4 review → themes that fit AI narrative not data

JTBD step invents jobs from model prior not study

Automation bias → no transcript re-read

Common mistakes

Treat AI themes as validated without sign-off

Batch all transcripts in one code pass

Remove dissent to "clean up" report

Let AI write recommendations as findings

Use public AI with identifiable participant quotes without consent review

PromptMake workflow

PromptMake /text: paste study context + phase number → get phase-specific prompt template.

Guest: 3/day. Registered: 5/day.

Build prompt chain library: P01 familiarization through P06 report — reuse per study.

FAQ

Can AI replace qualitative synthesis?

No. AI assists coding and drafting. Researcher owns theme validity, frequencies, and shareout.

Why Braun & Clarke as prompt chain?

Phases match how rigorous TA works — prevents skipping code and review steps that mega-prompts collapse.

What is frequency notation?

X of N participants (e.g. 8 of 15) — mandatory on every theme to fight automation bias.

Which model for interview synthesis?

Claude Opus 4.8 for coding and theme review. Flash only for first-pass highlights.

How many interviews per coding prompt?

One transcript per coding prompt. Merge at theme search phase.

How to handle contradictory participants?

Phase 4 negative cases + partial fit labels. Never smooth into single theme.

Can I connect to JTBD automatically?

AI drafts JTBD mapping from approved themes — researcher validates; tag speculative items HYPOTHESIS.

What about privacy?

Redact PII in transcripts. Use enterprise tools per org policy. Consent must cover AI-assisted analysis if applicable.

How long should quotes be?

Under 30 words typical. Longer only when necessary — mark and verify.

What is automation bias here?

Over-trusting fluent AI themes without verifying against raw data — mitigate with frequencies, quotes, and mandatory human gates.

Remote research synthesis adjustments

Remote sessions: code connection quality, screen-share friction, environmental distractions separately from product issues.

Frequency: X of N remote participants (label remote vs in-person if mixed study).

Stakeholder interview vs user interview

Stakeholder prompts use different codebook — business constraint, roadmap assumption, success metric.

Never merge stakeholder codes with user codes in theme search.

Tag STAKEHOLDER vs USER on every code row.

Contextual inquiry notes

Field notes include observation + environment context.

Code: observed_behavior | stated_preference | context_note.

Quote both what user did and what they said when they diverge.

Priority matrix prompt (post-themes)

INPUT: approved themes with X/N and business impact scores from team (human-entered).

TASK: Draft impact vs evidence matrix placement — not final priority.

OUTPUT: theme | X/N | suggested quadrant | rationale from quotes only.

PM moves quadrants — AI proposes from evidence text.

Negative findings report section

Required in phase 6:

SECTION: What we did not find — themes searched but below threshold (<3/N).

Prevents false certainty that study covered all possible pains.

Research plan linkage

Paste research questions from plan at top of phase 6 prompt.

Map each RQ to theme or NOT ADDRESSED with reason (sample, scope).

Closes loop on plan vs delivery for stakeholders.

Usability test synthesis prompt

TASK SYNTHESIS (not discovery interview — adjust phase 2 codes):

Code: task_success | task_failure | hesitation | workaround | quote.

Frequency: tasks failed by X of N participants on Task 3.

Link findings to task scenario text — not general product themes only.

Diary study synthesis

Entries are time-series per participant.

Code per entry with date. Theme search includes temporal pattern (e.g. frustration increases day 4–7).

Frequency: X of N had pattern — show timeline snippet per P ID.

Survey + interview hybrid

Quant survey scales + qual interviews:

Prompt: Cross-reference — theme (8/15 interviews) vs survey item mean. Label CONVERGENT or DIVERGENT.

Do not merge into single narrative without noting method difference.

Accessibility research notes

Code assistive tech mentions, barrier types (perceivable/operable/understandable/robust).

Quote AT users verbatim — do not sanitize disability descriptions.

Researcher validates respectful framing before shareout.

Service blueprint connection

APPROVED THEMES → map to journey stage (discover, onboard, use, support).

OUTPUT: stage | theme (X/N) | frontstage moment | backstage gap hypothesis.

Tag backstage gaps HYPOTHESIS — not observed unless ops interviews done.

Workshop readout prompt

INPUT: approved theme cards.

OUTPUT: 90-minute workshop agenda — affinity review, prioritization vote, JTBD framing exercise.

AI drafts agenda; researcher facilitates.

Research repository tagging

Export JSON codes with tags: study_id, date, product_area, method.

Enables cross-study theme search later — prompt includes STUDY METADATA block on every phase.

Inclusive language in codes

Rule: Codes describe behavior, not user identity (use "participant skipped verification" not "lazy user").

Researcher audits code labels before phase 3.

Related articles

Summarization prompts — preserving detail in long transcript summaries

Structured output prompting — JSON codebook schemas

Positive framing — verbatim vs paraphrase rules

Shareout slides should put frequency notation in the headline — "Checkout distrust (9/15)" — not buried in body text. Automation bias hides in polished themes when prevalence is hard to find.

When product asks what to build, separate findings from recommendations in every report prompt. AI-drafted roadmap items belong in a HYPOTHESIS section validated by researchers, not in the findings column.

Bottom line

UX synthesis prompts follow Braun & Clarke in six chained phases with human gates — not one-shot theme generation.

Frequency notation (X of N), verbatim quotes, automation bias awareness, and researcher sign-off are non-negotiable. Claude Opus 4.8 for coding and review. AI drafts; researchers validate.