01
Ingest & analysis
WorkingShot boundaries, a beat grid, variable-frame-rate normalisation, a word-level transcript, and scene-aware keyframe extraction that catches what a percentage-of-changed-pixels metric cannot — text changes, and subjects under 1% of the frame.
This is the stage that produces the candidate sets. Everything downstream depends on it.
02
Generation
Partly builtBrand-locked image and reel prompting built from a brand profile, so the same brief produces on-brand output twice. Routed to open-weight models that run on your own GPU first, where generation costs nothing per second.
Built and tested against an injected transport. No adapter has called a live vendor yet.
03
Edit
WorkingClaude proposes operations over MCP; the desktop app edits the same document directly. Both write the same format, so an instruction typed in the app and an instruction given to Claude produce the same result.
27 agent tools verified end to end. The desktop viewer is still being wired to the timeline.
04
Approve
WorkingTwo versions in, a readable change list out, with every operation carrying the reason it exists. Nothing renders as a deliverable without passing the gate, and the level is derived from the environment rather than requested by the caller.
A preview cannot be mistaken for a master, because the type system will not allow it.
05
Publish
Designed, not builtThe approved version hands off to publishing with its approval attached, so performance data comes back against the cut that produced it — and “which version performed better” becomes a query instead of an argument.
Designed and specified. The bridge is not built yet.