Keep Personas and Journey Maps Evidence-Linked

Workflow area: Research & Synthesis · Tier: Foundation · Hands-on time: ~30 minutes

A persona is only as trustworthy as the line between each claim and the research that produced it. Most personas break that line the day they're written: attributes get smoothed into a believable character, and six months later nobody can say which parts were evidence and which were vibes. AI makes the repair affordable — it can trace claims to sources at a scale no researcher has time for — but only if you force it to cite, and audit what it cites.

When to use — and when not to

• Use it after every research rounda persona refreshed quarterly from real studies stays a tool; one refreshed never becomes set dressing.
• Use it when inheriting artifacts you didn't makethe sweep tells you what's load-bearing evidence and what's decoration before you build on it.
• Use it before a big planning cycleroadmaps built on stale journey maps inherit their errors at 10x leverage.
• Skip it if you have no real research to link toAI cannot conjure evidence; with an empty repo, run pb-02 and pb-13 first, or label the artifact "assumption persona" honestly.
• Skip AI-generated persona filler entirelya synthesized stock photo biography is the exact anti-pattern this playbook exists to kill.

The recipe

1. 1.

   Gather the evidence base. Export your synthesized findings — themes reports from pb-13, usability findings, survey summaries — from Dovetail or wherever they live. Each needs a stable link or document name; the links are about to become load-bearing.

2. 2.

   Run the stale-claim sweep on the existing artifact. Paste the current persona or journey map text into Claude with your evidence:

   Copy

   You are auditing a persona/journey map against our actual research.
   
   Artifact: [PASTE PERSONA OR JOURNEY MAP TEXT]
   Evidence base: [PASTE FINDINGS SUMMARIES, EACH WITH ITS SOURCE NAME
   AND DATE]
   
   For every factual claim in the artifact, output one row:
   CLAIM | STATUS | EVIDENCE
   Where STATUS is one of:
   - SUPPORTED — cite the source name and the specific finding
   - CONTRADICTED — cite the source that contradicts it
   - UNSUPPORTED — no source in the evidence base addresses it
   Do not guess. If evidence is ambiguous, mark UNSUPPORTED and say why.
   

3. 3.

   Triage the audit table. Good output looks like this:

   Copy

   CLAIM: "Approvers prefer reviewing on mobile"
   STATUS: CONTRADICTED — Q2 interviews (6/8 participants): approve on
   phone = rubber-stamping; see themes report §3.
   
   CLAIM: "Month-end close is the peak-stress period"
   STATUS: SUPPORTED — Q1 diary study, days 28-31 entries; Q2
   interviews theme 1.
   
   CLAIM: "Age 34, drives a Subaru, loves podcasts"
   STATUS: UNSUPPORTED — no source addresses this.
   

   Contradicted claims get rewritten from the citing source. Unsupported claims get deleted or explicitly relabeled as assumptions — narrative filler doesn't earn a citation, so it goes.

4. 4.

   Regenerate the artifact with citations inline. Ask the model to rewrite the persona or journey stage-by-stage, keeping only supported claims, with the source name after each claim — (Q2 interviews, theme 1). In the final Figma or FigJam artifact, make each citation a real link to the source in your repo. In FigJam, AI can draft the journey-map skeleton; you attach the evidence links by hand because that's the part that must be true.

5. 5.

   Schedule the sweep. Put a quarterly calendar block on the artifact's owner: re-run step 2 with whatever research landed since. Add a visible "last verified" date to the artifact itself — an honesty timestamp readers can check before they trust it.

Make it yours

• Claude Project per product area: keep the evidence summaries and current personas in a Project; each new study drops in as it lands, and the sweep becomes a five-minute re-prompt instead of a re-assembly job.
• Dovetail highlights as the citation grain: if your repo tags highlights well, cite at highlight level instead of report level — claims linked to specific clips survive skeptical stakeholders better than claims linked to 30-page PDFs.
• NotebookLM for cross-study questions: load all findings docs as sources and ask "what do we know about onboarding friction across studies?" — its inline citations point back to specific sources, useful raw material for step 2's evidence base.

Quality check

• Every surviving claim has a working linkclick three at random; a dead link is an unsupported claim wearing a costume.
• Contradicted claims were actually rewrittennot softened into compatibility with the old narrative.
• The model didn't invent evidencespot-check that each cited source really contains the finding attributed to it; models misattribute across documents, which is this recipe's own hallucination risk.
• Demographic filler is gone or labeledanything not traceable to research is either deleted or visibly marked "assumption."
• The verified date is current and honesta fresh date on an unswept artifact is worse than no date.