Planning

I built the planning conversation around two things the creator can supply: what they think, and what makes the subject theirs. Research supplies the surrounding evidence, objections, limits, and context. That distinction shapes the interface. The creator can ask for more research, develop a take, or go straight to recording.

Underneath the conversation, the system assembles the structure the media pipeline needs. I added a check that compares the model’s extracted claim and personal connection with the creator’s actual answers. Whitespace and smart-quote differences are tolerated; reworded or invented passages are dropped. Missing material becomes another question or a request for research. The model cannot declare the plan ready when the required information is absent.

This lets the conversation stay direct while preserving a useful distinction between words the creator supplied and context the system researched.

Recording and processing

I connected the recording Studio to a saved version of the plan. The receiving service validates the material and stores the plan itself, so a recording can refer back to the brief without depending on the planning session remaining open. The creator does not have to copy their talking points between tools.

I also worked on the job system behind media processing. Each job is tied to its source recording, plan, settings, and the results it consumes. Repeated requests can reuse completed work, and concurrent requests share an execution. If curation changes the selected clips, the dependent work receives a new identity instead of silently reusing crops or captions from an older selection. Stored results are checked against the recording and plan before reuse.

Release

In the separate release service, I built checks around the exact version of a clip that was approved. Before a draft reaches a posting provider, the service checks its current approval, whether the clip has since been regenerated, and whether its media passed the required checks. An older decision cannot automatically authorize a new cut.

I connected publication to status reconciliation as well. A provider accepting a request leaves the publication pending; subsequent status updates establish whether it actually published. Scheduled items that remain overdue can be surfaced separately. These distinctions make the workflow more understandable: a finished render, an approved draft, and a published post each mean something specific.