Workflow

If a video composition is code, it can be versioned, parameterised, brand-gated, and re-rendered on demand. This project tests how far an agentic, code-as-video workflow can take learning and product video, and where it should still stop and hand back to a person.

The first experiment closed with a question : can an agent take structured product information and turn it into a short, repeatable video composition? This second experiment is the answer. The interesting part is not the seconds. It is what this workflow does to the way Learning Program Owners, Learning Designers, and Learning Developers traditionally split the work.

Can video production start to look more like a dev workflow, where the composition is code, the output is reproducible, and the agent can help move through the rough edges? Experiment 01 – a hello-world Editframe project, scaffolded, iterated, and rendered locally, with the failure points captured for the next pass.

While creating a blog post with Claude Code, I noticed the changes were on GitHub but not visible in my local files. The reason was git worktrees – a Git feature that lets Claude work in isolation without touching your local working directory. Here is what happened and what the trade-offs are.

Building the same workflow in Claude takes an hour; in Copilot it means Power Automate and three to five hours. That gap is real — and it’s architectural. Here is how Claude’s skill system maps to M365 Copilot, and what the comparison reveals about where each system was designed to be used.