Can an agent crew take a video from idea to publish?

Hypothesis

Video production is a pipeline of crafts — brief, concept, script, art direction, storyboard, generation, edit, publish. If each craft becomes a skill and an agent crew runs the pipeline with a human directing at each gate, one person should produce finished multi-scene video.

What we built

A creative stack of ten pipeline stages, each an agent skill: brief and trend research, concept, script, art direction, storyboard, then generation — frontier image models for keyframes, video models for motion with first-frame locking so multi-scene sequences stay visually continuous, generated music — and programmatic assembly into a final cut. The director (human) approves at every stage; nothing batches past a decision.

Learnings

• Multi-scene coherence is the craft: lock each scene's first frame to the previous scene's last, and character/style survive the cut. Prompt-only continuity does not.
• Step-by-step approval beats batch generation — one wrong art-direction choice compounds through every downstream shot.
• Keeping a live session-state document per project turned chaotic generation runs into a process a director can actually steer.

Log

• 2026-07-08 — Multi-scene chaining pattern (first-frame locking) proven across a full piece with generated music and programmatic assembly.
• 2026-06-20 — Ten-stage pipeline defined as individual agent skills.