AI Marketing Panel
Open source under the MIT licence. View the source
An installable kit that builds a panel of 5 to 15 researched customer composites and runs each one independently against your draft.
The problem
A commissioned focus group is a research project, so most small businesses skip testing and ship on instinct. The kit lets the owner or marketer pressure-test copy, pricing or emails first. The README states it is directional evidence, not a replacement for real customers.
How it works
You describe a decision and supply customer evidence (support tickets, sales calls, reviews). The kit builds an explicit panel with a long, evidence-labeled dossier per seat, runs each seat in its own context without seeing the others, synthesizes reactions while keeping dissent, and scores the artifact with owner-approved weights. Deterministic TypeScript validators (init-panel, validate-panel, prepare-run, validate-run, verify-install) run on Node 22.18+, and room.sh manages named rooms in plain bash. It runs at three capability tiers (subagents, shell-only, chat-only) and records the tier used in RUN.json.
Features
- Panel of 5 to 15 customer composites, one sourced dossier each
- Isolated per-seat runs (parallel with subagents, sequential in fresh contexts otherwise)
- Synthesis that preserves disagreement plus a weighted, scored verdict
- Three rewrites with retest, and a prediction/outcome ledger for calibration
- Deterministic validators and byte-for-byte installed-payload verification
- Install by dropping the kit zip into the Claude app or pointing Codex at the folder (verified 2026-08-12 per README)
How it was built
Version 1.0.0 shipped on 2026-07-16 as a portable skill for AI coding agents: an audience-neutral panel skill, workflows, deterministic tools and behaviour tests. It let a small business test copy against researched customer composites before real customers saw it.
The core rule was that the AI writes and a script judges. The model wrote seat dossiers and reactions, and TypeScript validators decided whether they had enough substance, cited evidence and correct score arithmetic.
Rooms arrived in 1.2.0 so the expensive part, researching the seats, happened once. A rerun then took minutes, and old runs were archived instead of deleted.
Several things broke on the way. The first Node floor was wrong because type stripping did not reach child processes, a version guard written in TypeScript could never run below that floor, and a stray Bun lockfile shipped after Bun was removed. Each was fixed in a patch release.
AI coding agents built the kit under Rico's direction, and release 1.2.3 was the fix pass from a cross-family review by a Claude model and a GPT model. Rico kept the outputs labelled as directional evidence, not a stand-in for real customers.
Credits
README says it is a portable implementation of the synthetic focus-group workflow from Corey Ganim's interview with Justin Brooke.