Focus Group — Synthetic Audience Pretesting

Pretest content against a panel of deeply-researched AI persona dossiers before a dollar or a send goes out the door. Based on Justin Brooke's predictive-wear framework; Brooke reports ~85-92% correlation with human focus groups when personas are high quality.

The iron law: personas > prompts. The accuracy comes from the 1,400-word dossiers, not the panel prompt. Never run a panel with thin personas.

Panel location

Keep one panel per business in your knowledge base, e.g. {KNOWLEDGE_BASE}/{business}/focus-group/. Each panel dir contains: an index note (roster, business context, scoring-weight overrides), personas/*.md (the dossiers), and runs/ (run logs). Record your panel locations in your CLAUDE.md so sessions can find them without asking.

Mode 1: RUN (default) — pretest content

Input needed: the content (pasted text, a file path, or a PDF — extract PDF text first with your PDF tooling) and which business it belongs to. If the business is obvious from context, don't ask. Identify the content type: ad (ads, social posts, emails, headlines) or offer (sales pages, launch emails, webinar scripts, anything asking for money) — offer mode adds the yes/no buy question. Separately, flag social mode when the content's primary goal is shares/forwards/organic reach rather than direct response — social mode adds the yes/no share question, the sharing-motivation requirement in the copywriter pass, and the social-mode scoring notes in the prediction engine (see references/sharing-psychology.md).

Social-mode decision rule (don't guess):

• Asset asks for money or a signup NOW (paid ad, sales email, landing page, launch email with a buy CTA) → conversion is primary. Social mode OFF unless the operator says the goal is reach.
• Organic and unpaid (social post, newsletter, video script for organic channels, lead magnet meant to be forwarded) → social mode ON by default.
• Both goals genuinely present (e.g., launch email you also want forwarded) → run BOTH modes; the copywriter splits the asset (share-worthy body, conversion close) rather than softening the money CTA.

Pipeline

1. Load the panel. Read the business's index note (roster + context + any scoring-weight overrides). List personas/*.md.
2. Fire the panel in parallel. Spawn one subagent per persona (all in ONE message so they run concurrently; batch in groups if the panel exceeds ~8). Each subagent gets: the full text of its ONE persona dossier file (tell it the path to read), the content under review, and the panel prompt from references/panel-prompt.md. Each returns the structured feedback block. Personas critique from their OWN life — never as marketing experts. Failure protocol: if a persona subagent errors, rerun it once; proceed only if ≥75% of the panel reported (e.g., 6 of 8) — otherwise abort and restart the round. Record actual responders vs. roster size in the run log.
3. Copywriter pass. Using references/copywriter-prompt.md: embody a world-class copywriter ("embody", never "pretend"), digest ALL panel feedback + the original, write 3 optimized variants in the business's voice (per the index note), formatted as an internal team email quoting key feedback.
4. Prediction engine. Scoring is done by a fresh subagent that did not write the variants (independence keeps the 85 gate honest). Give it: references/prediction-engine.md, all panel feedback, the business context + any weight overrides from the index note, and the variants. Default weights: Relatability 25 / Clarity 20 / Emotional 20 / Credibility 15 / CTA 20. Scores must anchor in panel data (see the "Who scores & anchoring" section of prediction-engine.md), not vibes. Verdicts: 85+ run, 75-84 test vs control, 65-74 revise, <65 reject.
5. Iterate. If no variant hits 85, feed the scores + weakest categories back through step 3 — and before round 2's copywriter pass, consult a persuasion-advisor layer if you run one (classic direct-response frameworks work well here: Eugene Schwartz's awareness/sophistication diagnosis for angle mismatches, Alex Hormozi's value equation for offer problems, and — in social mode — the NFX sharing motivations in references/sharing-psychology.md to diagnose why personas won't share). The panel says what missed; the advisor layer says why and how to fix it. Each round scores that round's 3 new variants, carrying the best prior variant forward as a comparison control. Max 2 rewrite rounds. After round 2, deliver the best variant with the engine's verdict as the ruling: 85+ ship; 75-84 offer it to the operator as a measured test against an existing control (their call); 65-74 stop — more wordsmithing is unlikely to close the gap, surface the weakest categories; <65 report that the offer/angle itself (not the wording) is the problem and route the offer redesign through your advisor layer before any further copy work.
6. Write the run log to focus-group/runs/YYYY-MM-DD-HH_MM-{slug}.md — see template below. This is non-optional; run history is how the engine gets calibrated against real results.
7. Report to the operator: panel highlights (2-3 verbatim persona quotes), the BUY tally for offers and the SHARE tally for social mode (note which personas SHOULD say no — wrong-fit "no"s are a feature; for shares, report WHO each yes-persona would send it to), the 3 variants, scores, and the recommendation.

Run log template

---
type: focus-group-run
project: {business-slug}
date: YYYY-MM-DD
---
# Run: {slug}
**Content type:** ad | offer  **Rounds:** N
**Panel:** {responded} of {roster} personas — BUY {X yes / Y no} (offer mode) · SHARE {X yes / Y no} (social mode)
## Key feedback
- {persona}: "{verbatim}"
## Scores
| Variant | Rel /25 | Clar /20 | Emo /20 | Cred /15 | CTA /20 | Total | Verdict |
## Shipped
{Which variant shipped, where, and — filled in later — how it actually performed}

Mode 2: SETUP — build a panel for a new business

Trigger: "focus-group setup {business}" or running against a business with no panel. Follow references/setup-mode.md: research the business from your knowledge base + web, then draft ~8 dossiers using references/persona-template.md (mix of struggling/successful, 1-2 deliberate wrong-fits), write the index note, and hand the panel to the operator for review — drafts are v1 until they've been sharpened with real customer knowledge.

Rules (hard-won, from our first panel build)

1. Personas > prompts. A lazy "pretend to be a parent" gets garbage. 1,400-word dossiers get the accuracy Brooke reports.
2. "Embody" beats "pretend you are" in every agent prompt.
3. Parallel over sequential — the panel has no cross-dependencies.
4. 85+ before running unmonitored; 75-84 only as a measured test against a control. Below 75 after two rounds = stop and be honest about why.
5. Wrong-fit "no"s are signal, not failure. Beginner offers should repel advanced personas and vice versa. Always report WHO said yes.
6. Never skip the run log. Prediction without calibration is astrology.
7. Proof/testimonials must match the audience segment (pro-athlete proof doesn't move a high-school parent).