Build evidence-backed behavioural personas
Create constrained behavioural persona drafts from coded evidence, then audit every attribute before using them to inform design decisions.
Why
The workflow problem
Teams need memorable user types, but unconstrained persona generation can turn stereotypes, invented demographics, and weakly supported attributes into design facts.
AI can draft candidate groupings and concise descriptions from a controlled schema. It can also amplify stereotyped patterns; the value is contingent on evidence tracing, prompt comparison, and human rejection of unsupported detail.
Concrete output
What you will produce
A behavioural persona set with needs, behaviours, evidence citations, confidence labels, excluded attributes, and a prompt audit log.
How
Run the workflow
Decide whether personas are warranted
The research lead confirms that the evidence contains meaningful behavioural variation relevant to a decision. Define a schema limited to behaviours, contexts, needs, constraints, evidence IDs, and confidence; exclude decorative demographics.
Generate bounded candidate groupings
Provide only the coded evidence and schema. Ask AI for competing groupings, each with traceable support and counterevidence, not polished characters.
Propose behavioural groupings with evidence
Work only from the supplied coded evidence. Persona schema: behaviour | context | need | constraint | evidence IDs | confidence | counterevidence Evidence pack: [insert] Propose up to [number] candidate behavioural groupings. For every attribute, cite evidence IDs and an exact excerpt. List attributes that cannot be supported and must remain absent. Do not add names, ages, gender, ethnicity, income, motivations, or biographies unless explicitly supplied and decision-relevant.Audit variants and verify attributes
Run a second, materially different prompt or human-only grouping exercise, then compare overlap and drift. Researchers verify every retained attribute against the corpus and document removals, disagreements, and attributes needing new research.
Publish only decision-useful personas
Design and product leads test whether each persona changes a real design or service decision. Publish citations and confidence with the persona; retire any profile that cannot be distinguished by evidence or use.
Human–AI partnership
Who contributes what
AI contribution
Candidate grouping, constrained prose drafting, and prompt-variant comparison. Output is a provisional representation, not a factual user model.
Human responsibility
Determine whether segmentation is warranted; govern the schema and data; verify attributes; identify stereotype drift; validate with further research; and decide whether the personas are useful.
Stop rule
Stop and discard the persona draft if an attribute lacks a traceable source, if pain points are assigned through stereotype, if fictional demographics begin to influence decisions, or if a prompt variant materially changes unsupported claims.
Check before use
Watch out
Personas can falsely imply stable, homogeneous populations. Keep evidence citations visible, distinguish segments from individuals, include counterexamples, and involve people with relevant lived or operational knowledge when representation risks are material.
Evidence base
What supports this workflow
Harvested Touchpoint material documents prompt A/B testing and failure cases in AI-assisted persona work, while NN/g warns that AI analysis needs human interpretation. This supports auditability as a safeguard, not a claim that generated personas improve outcomes generally.
- Chueng-Nainby, Gelshirani and Lee · Co-Designing Personas with AI
Harvested source documenting reproduced prompts, failure cases, and grounding practices for persona generation, including the importance of logging and comparing prompts.
- Maddie Brown and Kate Moran · Accelerating Research with AI
Establishes that AI-generated analysis requires human contextual interpretation and may miss or manufacture insights.
- Nielsen Norman Group · UX Research Cheat Sheet
Provides method-selection context: persona-like representations should remain connected to appropriate research rather than substitute for it.