Synthetic focus groups, run on your own box.
Focus groups are something a crew's marketing persona does, not a separate product. It puts its creative in front of a cohort of synthetic panelists, runs at least three rounds on your box, revises between them, and sends what survives to an approver as one card. Where your app is wired for it, that creative can then run as a real experiment.
The homepage headline earned its place: 56 → 72 → 78 over three rounds. One objection is still open.
AP-41 — Final creative for x-20260915-keep-the-slot (2 assets, cohort appointment-owners v1, score 56 → 78 over 3 rounds)Approving marks it ready to test. Nothing ships until a separate launch card.
A room built from a public library
Cohorts are built on klingbar.com from our panelist library: 197 invented people, 107 consumers and 90 business buyers, none of them real. Consumer panelists carry an Esri Tapestry segment, business panelists a 2022 NAICS code, and every one scores all 19 marketing channels.
A cohort is 6 to 16 of them, each worth one to three seats, with tags, a rationale, and a product-context block saying what you sell. Every save is a new version, and the box picks it up like any other prompt.
- shaped like
- Owners and front-desk managers of small appointment businesses
- members
- 7 panelists, 9 seats
- tags
owner·practice·services- context
- What the product does, what it costs, and who it is for
From a cohort to a rollout
The rounds run on your box under your Claude Code login. The experiment they open against is registered first and pinned to a Statsig layer, and the launch and rollout run where your app runs Statsig and reports conversions the way the persona expects.
-
The persona opens a focus group
Its
focus-grouptask takes a pre-registered experiment and the cohort whose tags match it, writes the creative for each arm, and brand-checks every piece before a panelist sees it. -
Each round runs on your box
One engine call per panelist, two at a time, on the persona's own Claude Code login. Each call loads no connectors, settings, or secrets, and runs in a fresh, empty directory holding only the stimulus. From round 2 a panelist also sees its own previous reaction, never another panelist's words.
-
A synthesis after every round
A separate call with a fixed rubric scores the round from 0 to 100, counts resonant and dissenting panelists from their own closing
STANCElines, and lists what landed, the objections, their phrases, and a revision brief. The persona revises and goes again: at least three rounds, at most five, and a stop rule says when. -
A card for the final creative
If the creative earned its place, one card names the cohort, its version, and the score across rounds. If not, the objection that sank it becomes a proposed learning.
-
A launch card, then a daily look
Approved, it merges the code, deploys, checks the revision now serving, and starts the experiment in Statsig, stopping at the first step that does not verify. Then the persona looks at every experiment in flight each day.
-
A decision, then a rollout card
A look can end in a winner, equivalent, futile, harm, or the exposure cap. A winner becomes a rollout card; everything else goes back to control.
Four kinds of card, and the rest runs on its own
Ideas, rounds, syntheses, daily looks, and the winner call run unattended, and report in #marketing when there is something to say. Stopping for harm needs no approval: the persona sends everyone back to control at once and says so.
The persona's prompt carries the rule that nothing launches, rolls out, or spends without an approved card. It works with the access you give it, so give it only what the job needs.
- final creative
- After the last round. Approving it ships nothing.
- launch
- Ship, verify, start. It also approves bounded daily re-weighting and the stops back to control.
- rollout
- The winner to 100%, the cleanup merged and deployed.
- spend
- Closed. A test that would need paid media ends as a question about budget.
Every round on the record
Three pages under each crew, and none has a control that changes an experiment: every change reaches you as a card.
Cohorts
The builder, each member linked to the library, every version, and whether the box applied it. After a decided experiment, what the cohort predicted sits beside what the market chose.
Focus groups
Each round's stimulus, questions, and synthesis, where each panelist stood in its own words, and every transcript.
Experiments
The pre-registration, immutable once written; each arm's creative source; a table of daily looks; and every decision with its card.
What a synthetic room is for
Synthetic panelists produce hypotheses, not evidence. Use a focus group to shape the question, find objections before a customer does, and cut weak variants early. The experiment is the evidence.
A cohort is shaped like an audience; it never stands in for one. It is a rationale and a handful of invented people, and the cohort pages say so.
klingbar.com does not add a marketing persona for you. The persona and its experiment engine were built against one stack: Statsig layers for the arms, and a ledger of conversions. Running the experiment it proposes needs the same wiring in your app: your repository as the persona's workspace, a Statsig project whose console key is sealed to your box, and a conversions source for the metric that decides. During early access we set up the marketing persona with you; it is not self-serve. The rounds run under your own Claude Code login, on your subscription. When synthetic respondents help, and when they do not.
Ask the room before you ask the market
Focus groups are free during early access, and we set up the marketing persona with you. Follow one from start to finish, meet the panelists, or read what a synthetic focus group can tell you.
Request early access