persona-@SLUG@

Tag the agent with a population_id, and it runs three passes: extracting cited observations per participant, clustering participants by shared goals and pain points, then synthesizing one persona per cluster with evidence citations. Every attribute traces back to a verified transcript line.

Each re-run creates a new draft version computed as MAX(version) + 1 by the write_persona_set tool, which supersedes the prior version in the same transaction. A re-run with identical personas is refused, and superseded versions remain linked for audit.

Roughly five interviews per population is the floor for meaningful clustering. Below that, the agent refuses to force personas, reports that the population needs more interviews, and flags it for human review.

No. Clustering uses only shared goals, pain points, and mental models — never demographics or raw cohort values. It does report cohort alignment diagnostically, stating whether emergent clusters line up with the raw cohort value.

The agent runs in emergent mode, clustering on whatever axes the data shows without goal-based filtering. All outputs — the structured record, Drive doc, and report — carry a goals_available: false flag and an explicit emergent-mode note.

It stops if the population_id is missing from the conversation_populations view or every row resolves to NULL, since that means no human has ruled on population mapping. It reports the gap rather than inferring membership from raw cohort values.