Content workflow
Plan a piece of content as a card on a board. Each column is a step; each step is either a deterministic tool or a bounded, iterative mini-chat; moving a card to the next column is an approval gate. The card is the durable home for everything (brief, prompts, images, copy), so chats never have to remember the whole history — nothing drifts, nothing is lost.
The pipeline
A step is a container, not one call
Within a column you go back and forth with the agent — using that step's skills and memory — until it lands. Advancing freezes the result onto the card and opens the next step's chat, seeded from the card. The next step reads the frozen output, not the transcript.
Fan-out to channels (repurposing)
Labels on the card target channels and pick the format. Each target spins a swimlane / variant card, so one concept becomes many channel-optimized posts.
Grounding = no hallucination
Copy is drafted from the card's own stored prompts + images + the venture brand kit (voice, styleguide, reference images), not from you re-pasting into a fresh chat. Deterministic steps (image generation, attach, schedule, post) use no model at all.
Design detail: workflows are shipped as hardcoded config (a
WorkflowDefper flow), the engine and board are generic, and the brand kit + card are the per-venture / per-piece data that layer in.