Test marketing copy against AI customer panels before it ships
If you write your own landing pages, emails, and pricing copy, you almost certainly ship without testing. Real focus groups cost roughly $20,000 and up and take weeks to organize. Most solo operators skip the test entirely and go on instinct. ai-marketing-panel is an open source kit that gives you a structured alternative in about ten minutes.
What It Does
The kit builds a panel of 5 to 15 customer composites drawn from evidence you already have: support tickets, sales calls, reviews, and the reasons people give when they say no. Each composite gets a detailed, sourced dossier. You feed the panel a specific decision (should this landing page go live, does this price hold up, will this email get opened) and each composite evaluates the artifact independently, without seeing what the others said.
You get separate reactions per composite, a synthesis that preserves disagreements rather than averaging them away, and a scored verdict you can push back on. The kit also produces three rewrites and retests when useful, and records predictions against actual outcomes for calibration over time.
The Honest Comparison
The kit’s own README is unusually direct about what you are and are not getting:
The honest part of the comparison is the last row, and it does not favor the kit: these are composites, not customers.
The $20,000 figure is described as an order-of-magnitude estimate that varies by scope and market, not a measured result. Per-run model costs vary too: a 15-seat run on a frontier model costs meaningfully more than a 5-seat run on a cheaper one. The kit recommends pricing your own run before quoting anyone else these numbers.
What the kit actually buys is the ability to run the test at all, at a cost where testing every artifact becomes a habit rather than a budget decision.
Harness Compatibility
Three harnesses are documented:
- Claude app: drag-and-drop the kit zip into the conversation. Verified working as of 2026-08-12.
- Codex: unzip to a folder, point Codex at it. Verified working as of 2026-08-12.
- claude.ai web: presumed zip upload, same as the app. Listed as unverified — never tested.
The kit runs at three tiers. T1 (Claude Code with subagent support) runs seats as parallel subagents. T2 (shell access, no subagent orchestration) runs seats sequentially in separate fresh contexts and is described as full-fidelity, just slower. T3 (chat-only, no shell) degrades to manual checklists.
️ Technical Requirements
The core scripts require only bash and coreutils, with no install step and compatibility back to Apple’s stock bash 3.2. The panel validators are TypeScript and require Node 22.18.0 or above. The kit notes this floor was raised from 22.6 at release 1.2.2 because Node only reads its TypeScript directly without a flag from 22.18.0 up. If you are on 22.6 to 22.17, run the preflight check first:
bash skills/synthetic-customer-panel/scripts/check-node.sh
Install current LTS from nodejs.org. On macOS with Homebrew: brew install node. Below the floor, room lifecycle still works but panel validation drops to T3 self-report mode.
The kit is MIT licensed, maintained by rTech Digital Consulting, and based on the synthetic focus-group workflow demonstrated in Corey Ganim’s interview with Justin Brooke. Hosted install and tuning are available from the maintainer.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
