Simulate campaign (synthetic reader panel)

Show each creative to each persona, have the persona react in first person with scores and objections, then rank and fix. The personas live in personas/; each is a self-contained prompt ("You ARE ...").

What a persona panel is good for, and what it isn't

Good at: catching copy defects. A headline nobody would stop for, a pitch that smells like a sales funnel, a number that reads as machine-guessed, a deliverable written to the wrong reader (owner advice in a broker brief), an objection the page never answers.

Bad at: predicting who will reply or buy. In a backtest of persona panels against real outreach outcomes, panel scores ranked actual replies at about chance (AUC ~0.5). Treat scores as a defect detector, never a forecast, and never report "personas say 70% will click".

Synthetic personas also drift toward agreement. Counter it: the persona file says what makes them bounce; ask for objections before scores; and run each persona as a separate call or subagent that sees only its persona file and the creative, never your reasoning or the other personas' answers. (For a fuller anti-sycophancy protocol on a live product, see role-litmus.)

The panel

File	Who	Use for
maria.md	struggling full-service owner, cash-tight, scrolls late at night	ads, landing pages
tony.md	established Italian owner, skeptical of tech	ads, landing pages
dave.md	overwhelmed bar & grill owner, no bandwidth	ads, landing pages
yara.md	cost-squeezed Mediterranean owner, needs proof	ads, landing pages
jake.md	data-literate cafe owner, needs depth	"is this too basic?"
acquirer.md	PE / roll-up buyer (home services)	reader-QA on a deal feed or target list
broker.md	commercial real estate broker	reader-QA on a prospecting dossier or brief

Write your own from the same shape: who they are, their numbers, their psychology (pain, pride, skepticism, when and where they read, what stops the scroll, what makes them click, bounce and buy), then the scoring protocol. Adjust prices in the files to your offer.

Run it

1. Collect the creatives. For ads: headline, primary text, image description, CTA, and the landing page text (fetch the live page so the persona reads what ships). For a deliverable: the draft itself.
2. Match personas to creatives. Each sector's ad goes to its sector persona; run a cross-test (every persona x every ad) when the targeting is broad.
3. One isolated call per (persona, creative): the persona file as the system prompt, the creative as the message, and an instruction to answer in this JSON:

   {"scroll": 0, "click": 0, "trust": 0, "buy": 0,
    "monologue": "...", "objections": ["..."], "one_change": "...", "verdict": "SHIP|REFRAME|null"}
   

4. Rank variations by mean score; list the objections that recur across personas.

Reading the scores

Score	Meaning
8-10	strong
6-7	workable with the persona's one change
4-5	weak: read the objections, rewrite
0-3	dead for this audience

• Scroll is the gatekeeper. Under 5 and nothing else matters: fix the headline first.
• High scroll + low buy = the ad hooks and the landing page doesn't close.
• Reader-QA personas end on SHIP or REFRAME. An audience-test failure (owner-facing advice in a broker or buyer document) is REFRAME regardless of other scores.
• Objections that two or more personas raise are the real fixes; a single persona's objection is a curiosity.

Iterate

Fix the recurring objections, re-run, compare. When the top variants clear 7 across their target personas, the copy is free of the defects a panel can see. Then measure the real thing with real traffic.