Insta

The classify stage sends the transcript, the video description, the caption, my note and the numbers to a model. It suggests a pillar, up to three formats from the shared catalog (matched only within the row's medium: reel, carousel or still), a hook and a verdict of adopt, study or skip, each with a confidence. The result is stored as a suggestion. When I sort the row, my choice is compared with the suggestion field by field and stored, and an eval script reports how often we agreed on pillar, format and hook. The one thing it files on its own is a high-confidence internal pillar: a post about Instagram itself, not content to imitate.

Every time I sort a suggested row, my choice is logged against what the classifier said. After ten, a distill round runs in the background: the model reads them against its own suggestions and proposes up to three one-sentence rules, each backed by at least two rows where I disagreed. Gates drop near-duplicates, handles, ids and rules over 320 characters. Each survivor is replayed on the same ten rows and goes live after a 48-hour grace window only if agreement got strictly better with no conflict flagged; a tie waits for me. A reason typed in the sort panel becomes a rule at once. A rule applied fifty times without helping is flagged for retirement, never retired on its own.

Research is one pool shared by every brand; only the idea stage knows which project it works for. Once per project, the model reads the row against that project's brief, its own cards (the ground truth for its taste) and its idea rules, and decides whether the container is relevant. If so, it picks a format and hook from the catalog, writes a one-line idea in the project's world and a filming recommendation, and up to six shots with who is in each and who must be there. It never writes a hook line, script or caption. In the Create tab I edit it into a card, or dismiss it; a typed why becomes a project rule.