TL;DR
Feed raw transcripts from 8 to 12 interviews into a long-context model, and ask for a 1-page persona organized by Jobs-to-be-Done, with every trait backed by a verbatim quote and a required “what does not matter” section. The 2026 best practice (per Nielsen Norman Group and User Interviews) is to drop demographics in favor of role, behavioral signals, and decision criteria. Use Claude Opus 4.7 or Sonnet 4.6, Gemini 3.1 Pro, or ChatGPT (GPT-5.5) — all carry a 1M-token context as of June 2026, so 8 to 12 full transcripts fit in one prompt. The copy-ready prompt is below.
The task
You wrapped 8 customer discovery interviews over the last 3 weeks. You have a folder of recordings, partial transcripts, and a Notion page of scattered quotes. Your PM lead wants a “persona” document by Friday so the design and copy teams have a shared mental model. You also know that the most common failure here is producing the “Alex, 32, marketing manager, drinks oat milk” template — a stock-photo persona that lets the team pretend they know the user without anchoring any actual product decision. You want a 1-page persona built from real quotes, organized by Jobs-to-be-Done, with an honest “what does not matter” section that protects the team from inventing demographic noise.
Where AI helps — and where it does not
AI is good at clustering verbatim quotes from interviews into recurring patterns, mapping observed behaviors to Jobs-to-be-Done framing, and producing the “what doesn’t matter” section that most personas skip. It can also pattern-match decision criteria across interviewees and surface the consistent triggers — what’s happening in a user’s life when they go looking for your product.
What AI cannot do: invent traits not present in your interviews. If you ran 2 interviews and ask for a persona, AI will fabricate plausible-sounding details that aren’t grounded, and those fabrications get treated as data by the team. JTBD researchers generally cite 8 to 12 interviews before patterns are stable enough to act on (User Interviews, 2026); treat 5 as a hard floor and below it do more interviews first. AI also cannot tell you which segment is most worth optimizing for. That’s strategy, not pattern-matching.
A specific failure mode: AI tends to over-stuff personas with demographic details (age, location, job title) even when interviews showed those traits didn’t predict behavior. Tell it explicitly: “remove any trait that does not have a verbatim quote backing it, and explicitly include a ‘what does not matter’ section listing demographics that did NOT predict behavior.”
Which model and tools (June 2026)
The whole job is reading long transcripts and clustering quotes, so context window and instruction-following matter more than benchmark scores. Eight 45-minute interviews run roughly 60,000 to 90,000 words of raw transcript, which is well inside a 1M-token window. All four mainstream options below can hold the full set in one prompt as of June 2026.
| Tool | Plan to use | Context | Why for this job |
|---|---|---|---|
| Claude Opus 4.7 / Sonnet 4.6 | Pro $20/mo | 1M tokens | Best at holding the “no quote, no trait” rule across a long synthesis; Sonnet 4.6 is the cheaper workhorse |
| Gemini 3.1 Pro | Google AI Pro $19.99/mo | 1M tokens | Strong on long-document recall; handy if your transcripts already live in Google Docs |
| ChatGPT (GPT-5.5) | Plus $20/mo | ~320 pages in-app | Fine for 8 transcripts; the full 1M context needs the $200 Pro plan |
For the upstream transcription step, a dedicated note-taker is cheaper than re-recording: as of June 2026, Otter.ai Pro is about $8.33/seat/month and Fireflies.ai Pro is $10/seat/month, with Fireflies’ free tier covering 800 minutes against Otter’s 300. Export the verbatim transcript (not the auto-summary) and paste that into the model, so the persona is built on real words rather than a tool’s paraphrase.
What to feed the AI
- Detailed notes or transcripts from 5+ user interviews (the more verbatim, the better)
- The specific product decision the persona is meant to inform (positioning, onboarding flow, pricing tier design, ad targeting)
- Any demographic or firmographic facts that you observed actually changing behavior across interviews
- Demographics you suspect are noise (intuitive but unsupported by interview data)
- Your current draft persona, if you have one — so the model can diff against the actual data
- The 2-3 quotes from interviews that already caught your attention as defining
- The alternative products users mentioned they’d consider (these are decision-criteria signal)
- The trigger event — what was happening in their week when they went looking? (very predictive, often skipped)
Copy-ready prompt
Build a user persona from these interviews.
Interview notes or transcripts: {paste 5+ interviews}
The product decision this persona must inform: {paste}
Demographics that I observed changed behavior across interviews: {paste or "none confirmed yet"}
Demographics I suspect are noise: {paste}
The 2-3 verbatim quotes that already caught my attention: {paste}
Alternative products interviewees mentioned: {paste}
The trigger event — what was happening in their life when they went looking: {paste if observed}
Return a 1-page persona:
1) Persona name — descriptive, not a human name. Use "Solo Marketer in Mid-Growth," not "Sarah." A descriptive name forces the team to think behavior, not faces.
2) Trigger — the situation in their life when they go looking for a product like ours. Backed by quotes from at least 2 interviewees.
3) Jobs to be done — 3 jobs, ranked by frequency in interviews. Each job has a verbatim quote (in quotation marks).
4) Decision criteria — what makes them say yes vs. no. 2-3 yes-triggers and 2-3 no-triggers, each with a quote.
5) Alternatives — what they currently use or would consider instead. Lead with the alternative most mentioned.
6) What does NOT matter — explicitly list 3-5 demographics or traits that did NOT predict behavior across interviews. This section is required, not optional.
Rules:
- Every trait in the persona must be backed by a verbatim quote from at least 1 interviewee. No quote, no trait.
- Patterns must appear in at least 3 of the interviews to count as persona-level (avoid one-quote anomalies).
- The "what does not matter" section is the trust signal of the whole document; if you cannot fill it, the data is too thin.
- Total length under 400 words. One page.Shorter variant — quick segment hypothesis
From these 3 interviews, generate a hypothesis for a single user segment and 3 questions I should ask in my next 3 interviews to confirm or reject the hypothesis. Skip the full persona. {paste interviews}Sample output
A useful Jobs-to-be-Done entry: “Job 1 (high frequency, mentioned in 6/8 interviews): ‘Show me what changed in the last 7 days across my team’s KPIs without me having to build dashboards.’ Verbatim: ‘I just want a Monday morning email that says here’s what moved and here’s what to look at — I don’t need another tool to log into.’”
A useful “what doesn’t matter” entry: “Age does not predict behavior here — interviewees ranged from 24 to 51 and the decision criteria were the same. Industry does — SaaS marketers focused on retention metrics; agency marketers focused on client-reporting cadence. Company size also did not predict — solo founders and 200-person teams had identical ‘don’t make me log into another tool’ criteria.”
A useful trigger statement: “Trigger: end-of-quarter board prep period, when they need to assemble metrics from 3-4 different tools and have spent the previous quarter promising ‘next quarter I’ll have better visibility.’ Verbatim from 3 interviews: ‘I always say I’m going to figure this out, but then the quarter ends and I’m pulling data manually again.’”
A useful decision-criteria block: “Yes-triggers: (1) Works with our existing stack (no new logins). (2) Output is a written summary, not a dashboard. (3) Can be customized in under 10 minutes. No-triggers: (1) Requires a setup call with sales. (2) Pricing requires a custom quote. (3) Branded as ‘AI-powered’ without showing what the AI actually does.”
How to refine
- Drop any trait without a quote: “Re-read the persona. Any trait, behavior, or demographic without a verbatim quote in quotation marks gets removed or flagged as ‘hypothesis — confirm in next round.’ This rule is the difference between persona and fiction.”
- Demand the ‘what doesn’t matter’ section: “If the ‘what does not matter’ section is empty, you have not done the work. Go back to the interviews and identify at least 3 traits that varied across interviewees without changing decisions. That’s the section that protects the team from inventing noise.”
- Map jobs to frequency: “For each Job-to-be-Done, cite how many of the interviews mentioned it (e.g., ‘6/8 interviews’). Jobs mentioned in fewer than 3 interviews go to the ‘hypothesis — needs more data’ list, not the main persona.”
- Use descriptive persona names: “Replace any human name (Sarah, Alex) with a descriptive label tied to behavior (Solo Marketer in Mid-Growth, B2B PMM Pre-IPO). Human names anchor the team on a face; descriptive names anchor on behavior.”
- Tie the persona to one decision: “Re-read your output through the lens of the product decision this is meant to inform. If a section doesn’t help anyone make that decision better, cut it. Personas exist to drive decisions, not to be archived.”
Common mistakes
- Inventing demographic details to make the persona feel “real” — backfires in product decisions because the team starts optimizing for traits that don’t predict behavior
- One persona per interviewee — you want the cross-interview patterns; if you find yourself with 8 personas from 8 interviews, you have not synthesized
- Including demographics that did not predict any behavior — they mislead the team into making decisions based on age, location, or job title when the actual driver was something else
- No “what does not matter” section — its absence is a tell that you haven’t done the synthesis work; the team will fill in noise on its own
- Stopping at 2-3 interviews — below 5 the patterns are too noisy and AI will fabricate to fill the gap; 8 to 12 is where JTBD patterns actually stabilize, so do more interviews first
- Using a human name with a stock photo — anchors the team on a face and a back-story that aren’t grounded in data; use descriptive labels instead
- Treating the persona as static — interviews 2 months from now will refine it; the persona is a living document tied to ongoing discovery, not a launch deliverable
- Building 5 personas before validating one — the second persona only earns its place when there is evidence that segment makes different decisions for different reasons
FAQ
- How many personas should I build?: Start with one. Add a second only when you have direct evidence that a second segment makes different decisions for different reasons — not just different demographics. Most products that ship “5 personas” are using personas as marketing artifacts, not decision tools.
- Do I need a profile picture for the persona?: Optional, and often actively harmful. A stock photo anchors the team on a face and an inferred back-story neither of which came from data. Use an icon, an abstract illustration, or just text.
- How often should I refresh the persona?: Every time you do a batch of 5+ new interviews, or when you launch a feature that changes the user mix. Persona drift is invisible until a sales call reveals you’ve been targeting an outdated profile for 6 months.
- What if my interviews contradict each other?: Contradictions are signal. Either you have 2 segments (build 2 personas, with explicit decision differences), or the disagreement is on a demographic that’s noise (move it to “what does not matter”). Don’t average contradictions into a vague persona.
- The model keeps inventing details — what changes?: Add: “Every trait must be backed by a verbatim quote. If you cannot quote it, do not include it. Mark any inferred patterns as ‘hypothesis — confirm in next round’ in a separate section. The persona’s value is in its grounding; fabricated traits make the document misleading.”
- Which model handles 8 transcripts best?: Any 1M-token model works as of June 2026: Claude Opus 4.7 or Sonnet 4.6, Gemini 3.1 Pro, or ChatGPT on GPT-5.5. On ChatGPT Plus the in-app window is roughly 320 pages, which still holds 8 interviews; the full 1M context is only on the $200 Pro tier. Claude tends to be the most disciplined about the “no quote, no trait” rule across a long document, so it’s the default pick here.
- Should the persona include demographics at all?: Only the ones that changed behavior. The 2026 norm is to lead with role, Jobs-to-be-Done, and decision criteria, and to push noise-level demographics into the “what does not matter” section rather than the headline. See NN/G on personas vs. JTBD.
Related
- AI customer discovery questions
- AI MVP scope
- User Persona Prompts for Personas That Drive Decisions
- AI Interprets Survey Results
- AI Positioning Statement
Tags: #AI writing #Product #Workflow #Persona