I used Codex to turn a CSV of customer reviews into a pretty deep persona research system.
A few friends have tried it too and liked it, so figured I’d share the workflow.
Super simple:
1/ First, organize the reviews + quantify the evidence.
Have Codex pull out things like:
→ Pain points
→ Motivations
→ Desires
→ Objections
→ Customer language
→ etc.
Then count how many reviews actually support each one.
This part matters a lot.
If Codex says 37 customers mentioned a pain point, that number needs to actually be 37.
AI can’t get creative with the math lol.
Verify the counts yourself, or at minimum spot-check a few of them before moving on.
2/ Then assemble the research team.
I spawned 5 different subagents:
→ Customer Researcher
→ Copywriter
→ Paid Social Creative Strategist
→ Organic Social Strategist
→ Media Buyer
This was me just trying to force different POVs into the process.
The agents individually used the research to build 5 different personas.
Each persona got 10 specific “why now?” purchase triggers.
5 personas × 10 triggers = 50 triggers total.
Across all 5 agents, that gave me 250 research-backed purchasing triggers to work through.
I was pretty strict about what could actually make the cut.
Every trigger needed:
At least 1 real customer review, with the specific review cited
At least 1 external customer signal, usually ended up being a quote from Reddit
If there wasn’t good evidence for it, I didn’t want it in the doc.
Pretty simple.
3/ Last step: make the agents review the work against each other.
Nothing really fancy here.
I just wanted all five POVs to argue with each other until they landed on a strong conclusion.
I kept that process going until all of the agents agreed (with each other) on one final persona document.
That was probably the most useful part.
You’re having a bunch of different specialists build, critique, revise, and eventually agree on the same answer.
The final doc had:
→ 5 personas
→ 50 purchase triggers
→ Customer language
→ Objections
→ Required proof
→ “Why now?” moments
→ Evidence behind each trigger
And you can score those triggers by evidence, urgency, emotion, creative potential, scaling potential, etc.
So you end up with a research-backed bank of ideas for:
Ads
Landing pages
Emails
Offers
Creative briefs
Site messaging
Upload → quantify → assemble the agents → research → argue → agree.
Pretty sick stuff.
Comment how it works out for you if you try it out!
Peace and Love,
Travis