An AI focus group, one panelist at a time
In Klingbar, an AI focus group is a series of rounds that a crew's marketing persona runs on your box. Each round is one model call per panelist, on a Claude Code login you own, and every panelist answers alone. After each round a single synthesis scores it from 0 to 100, and the persona revises the creative before the next.
Every panelist answers in its own call
Put a dozen model-played respondents in one conversation and they start to agree with each other, so nobody here is in a conversation. Each panelist gets its own call, its own document as the system prompt, and a fresh, empty working directory holding only that round's stimulus.
The call loads no connectors, no settings or plugins, and no secrets, and reads outside the stimulus are denied. The persona's standing rules and learnings stay out of it, and so does every other panelist's answer. At most two calls run at once, each under a time limit the cohort sets, 180 seconds by default.
- system prompt
- That panelist's document, exactly as the cohort rendered it
- questions
- The round's fixed questions
- stimulus/
- The copy, the context, the page or email source, and an image when one could be rendered
- from round 2
- Its own previous reaction, word for word, and nobody else's
- not loaded
- Connectors, settings, plugins, secrets, standing rules, learnings, other panelists
The same questions, every round
Five questions stay fixed across rounds, which is what makes round 2 comparable with round 1: what did you think this was in the first ten seconds; what is it promising, and would you believe it; what in it is for someone like you; which line would you keep and which would you cut; and would you act, and what would stop you. The persona may add at most two of its own. Answers run 250 to 500 words, in the first person.
Every answer ends with closing lines in a fixed vocabulary: STANCE (positive, mixed, or negative) and WOULD_ACT (yes, no, or unsure). From round 2 the panelist also reads its own last answer and says whether the revision dealt with what it raised: ADDRESSED (yes, partly, or no). A line that is missing or off-vocabulary is recorded as missing, never guessed from the prose.
STANCE: mixed WOULD_ACT: unsure ADDRESSED: partly
An illustration, abridged. Real answers are longer.
What happens in a round
The persona assembles the stimulus
Copy, context, and the page or email source go into the round's stimulus directory. When no image can be rendered, the questions say so plainly and the round records that panelists reacted to copy alone.
Each panelist answers
One call per member, as above. Below two thirds answering, the persona reports the round as blocked rather than grading a hollow one. A usage limit on the login stops the round at once and names the login.
One synthesis reads the round
A separate call, on the cohort's synthesis seat, reads every answer against a fixed rubric: what landed, the objections, the phrases panelists used unprompted, and a brief for the next revision. The persona that wrote the creative does not grade it.
The persona revises
It writes down what changed, which objection each change answers, and which line it kept because it landed. The revised creative becomes the next round's stimulus.
Counted from the panelists, not from the summary
A model summarizing a dozen answers can report agreement they do not contain, so the synthesis supplies the reading and the closing lines supply the numbers. The score is positive seats plus half the mixed ones, over the seats that answered, times 100; a member holds one to three seats.
A panelist whose call timed out is left out of the count, not counted as a dislike. The rubric's own numbers are kept, and the page shows them only where they differ. klingbar.com records one synthesis per round and never re-grades one.
- answered
- 11 of 12 — one call timed out and is not counted
- stance
- positive 6 · mixed 3 · negative 2, from the panelists' own
STANCElines - addressed
- yes 5 · partly 4 · no 2
- landed
- Showing the entries it was unsure about
- objections
- Nothing about tax time; the price appears only after signup
- their language
- "what it would get wrong", "my accountant"
At least three rounds, at most five
After round 3 the group stops when the score rose by less than 5 points, or when dissent is down to a quarter of the seats. Otherwise it runs a fourth round, and never more than the cohort's cap of five at most. A score that falls two rounds running closes the group as not earned. To earn a card, the final round needs a score of at least 60 with dissent at no more than a third of the seats; the final creative set then goes to an approver, naming the cohort and its version.
A score of 68 means the panelists leaned positive on that draft, and nothing more. What a synthetic focus group can and cannot tell you sets out the limits, and the record on klingbar.com keeps every transcript behind the number.
Put your next draft through three rounds
Klingbar is free during early access. You bring the box and a Claude Code login, and we set up the marketing persona with you; the rounds run there, and the record is on klingbar.com.
Request early access