Captions written from the photos, approved once, published to every network

When a content item is ready, the caption scenario reads it, iterates its images and aggregates them into one set. A router with a route per network sends the images to OpenAI Vision, has a completion write that network's caption, and writes it to that network's column on the item, so the approver sees every version side by side.

Each network's route starts with a short sleep, staggered from one to five seconds, so the parallel calls do not reach the model's rate limit at once. Writes back to the board carry their own error handlers.

A daily scenario lists the posts scheduled for that day, fetches each one's media and attaches it to the item, and moves the item into the ready group. Posting is a separate step, so the schedule can change without touching the publisher.

The publisher fires on an approved post, and a router sends it down one route per network: a photo post, a carousel built by iterating the images, a pin, and a company-page image post, for which the file is fetched first because that module needs it. Each route ends by writing the post's status or link back to the item. A second build of the publisher, on an Airtable base, checks for approved posts every two days and posts to three networks.

Each network has its own error handling: an ignore handler where a failure should not stop the run, and a break with automatic retry on one network's route. One network being down never blocks the others, and each record shows exactly where it went out.