Most personas are demographic posters — “Marketing Mary, 32, urban, loves coffee” — that nobody opens twice. They drive zero decisions because they never say what the buyer does today, what they wish they could do, what disqualifies them, or what they Google at 11pm when they’re stuck. The 13 prompts below force behavioral specificity, real-quote anchoring, jobs-to-be-done links, and explicit anti-personas, so the output actually shows up in product, marketing, and pricing calls. Pair them with jobs-to-be-done prompts for the upstream framing.
TL;DR
- Feed the model real evidence (interview notes, sales transcripts, support tickets) — a persona built from a blank prompt is a proto-persona: an assumption, fine for kickoff, never for product requirements.
- Demand behavior over demographics: what they do today, what they wish they could do, decision criteria, and a disqualifier.
- Always generate an anti-persona (Prompt 3). The people who look like a fit but churn cost you the most.
- For long transcripts, use a 1M-token model: Claude Opus 4.7 / Sonnet 4.6 or Gemini 3.1 Pro all hold ~750k words at once (June 2026), so you can paste a full multi-hour transcript set rather than chunking.
- Re-run Prompt 12 every quarter. Markets shift; a two-year-old persona quietly stops being true.
Best for
- Pre-MVP product discovery
- Marketing and positioning work
- Designing onboarding flows
- Sales enablement
- Sharpening go-to-market
How to get a usable persona out of these
Personas earn their keep when they come from qualitative research — real interviews and calls — not from a clever prompt alone. So the workflow is: gather raw evidence first, then use these prompts to synthesize it. The model is a pattern-finder, not a substitute for talking to customers.
Two practical notes before you start:
- Replace every
[bracketed placeholder]with your own text. Where a prompt says[paste interview notes], paste the actual notes — the more raw quotes the better. - Pick the model by input size. Short briefs work anywhere. For Prompts 2, 7, and 12 (transcripts, quote banks, full doc updates), use a long-context model so nothing gets truncated. As of June 2026, Claude Opus 4.7, Claude Sonnet 4.6, and Gemini 3.1 Pro all ship a 1M-token window standard; ChatGPT Plus reads roughly 320 pages in-app, with the full 1M window reserved for the $200 Pro tier — so on Plus, chunk anything over ~150k words.
1. Behavior-led persona
Build a persona for [product]. Inputs: [paste interview notes / survey data].
Output: a behavioral name (not "Marketing Mary"), a 1-line context,
what they do today, what they wish they could do, their decision criteria,
and what disqualifies them as a buyer.2. Persona from sales calls
Below are 5 sales-call transcripts. Extract the underlying persona:
shared problem framing, shared vocabulary, shared decision triggers.
Avoid demographic mush.
[paste calls]3. Anti-persona generator
For [product], define 3 anti-personas: people who will look like a fit
but will churn or never buy. For each, give: why they look like a fit,
why they actually are not, and the earliest signal to disqualify them.4. JTBD-linked persona
Build a JTBD-linked persona for [product]. Output: the job they hire
[product] for, the situation that triggers the hire, the alternatives
they considered, and the trade-off they accept.5. Buying-committee persona (B2B)
For my B2B product, build the buying-committee personas: economic buyer,
technical evaluator, end user, and blocker. For each: their core concern,
what wins them over, and what loses them.6. Persona shift over time
My early customers were [persona A]. New customers feel different.
Inputs: [paste new-customer notes]. Identify the shifting persona and
what about our product attracted the new group.7. Persona quote bank
Below are 20 customer-interview quotes. Cluster them into 3 personas.
For each persona, give 5 verbatim quotes that capture their voice.
These quotes go straight into landing-page copy.
[paste quotes]8. Persona-to-positioning bridge
For persona [paste persona], write the positioning statement:
"For [persona], who [problem], [product] is a [category] that [benefit];
unlike [alternative], we [differentiator]." Then give 3 variants.9. Persona x channel map
For my 3 personas [paste], map: where they hang out, what they Google,
what content they consume, and what triggers them to seek a solution.
Use this to plan channels.10. Persona-pricing fit
For each of my 3 personas, evaluate fit at each pricing tier.
Output: which tier matches, the willingness-to-pay signal,
the objection at this tier, and what would justify moving up.11. Persona feature-prioritization filter
For persona [paste], filter my backlog: [paste 20 features]. For each,
mark: high-priority for this persona / nice-to-have / waste-of-time.
Justify each call from their JTBD.12. Persona update from new evidence
Below is my current persona doc, followed by 10 new customer interviews.
Update the persona: what to add, what to remove, what changes the
confidence level. Mark anything that is now contradicted.
[paste both]13. Persona pressure-test
Below is my draft persona. Pressure-test it on four axes:
(a) is it behaviorally specific, (b) is it grounded in real data,
(c) is it actionable, (d) does it have a meaningful anti-persona.
Score each 1-5 and propose fixes.
[paste]Common mistakes
- Demographic-only personas (“32, urban, marketing”) — no decision can be made from this.
- Personas not grounded in real interviews or data — vibes-based personas drive vibes-based features.
- No anti-persona — every user looks like a fit, so churn surprises everyone.
- A persona doc no PM, marketer, or designer opens again after kickoff.
- The same persona used unchanged for 2+ years while the market shifted around you.
- Cute persona names (“Marketing Mary,” “Founder Frank”) instead of behavioral names that carry information.
FAQ
What’s the difference between a persona and a proto-persona? A persona is built from qualitative research — real interviews, calls, support data. A proto-persona is built from your team’s assumptions. Nielsen Norman Group’s guidance is blunt about it: a proto-persona is fine for guiding the next round of research, but should never be the basis for product requirements. These prompts produce a proto-persona when you feed them assumptions and a real persona when you feed them evidence.
Personas or jobs-to-be-done — which should I use? Both, and they answer different questions. JTBD describes the outcome the customer wants; a persona describes who wants it and how they behave on the way there. The cleanest workflow is JTBD first for the framing, then a persona to make it concrete (Prompt 4 links the two). See our jobs-to-be-done prompts.
Which AI model should I use for the transcript-heavy prompts? Any frontier model handles a short brief. For pasting full transcripts or a 20-quote bank, pick a model whose context window holds the whole input: as of June 2026, Claude Opus 4.7, Claude Sonnet 4.6, and Gemini 3.1 Pro each offer a 1M-token window (~750k words) standard. On ChatGPT Plus the in-app limit is roughly 320 pages, so chunk larger inputs or move up to the $200 Pro tier for the full 1M window.
How many personas should I end up with? For most early-stage products, two or three is the working maximum. More than that and no team can hold them in their head, and the doc stops driving decisions. Prompt 7 deliberately clusters into 3.
How often should I update a persona? Treat it as living. Re-run Prompt 12 against new interviews at least quarterly, and immediately after any pricing change, new-segment win streak, or pivot. Personas decay because markets, products, and the competitive set all move.