Early in our AI Labs work, we ran into a problem that felt embarrassing to admit: we could not reproduce our own results. A sequence that had processed beautifully three weeks earlier looked wrong when we ran the same pipeline again. The model was the same. The plates were the same. The prompt was approximately the same. Approximately is not good enough in production.
Prompts are code
Every prompt that feeds an AI step in our pipelines is now version-controlled in Git alongside the pipeline code it belongs to. Prompts have semantic version numbers. Changes require review from at least one senior artist on the relevant show — not just an engineer.
The review process
Prompt changes go through three stages: automated regression tests against a reference batch of frames, visual review by the artist who owns the sequence, and sign-off from the VFX supervisor. For hero sequences the supervisor sees a side-by-side comparison on actual production frames before anything merges.