Case study
Prompts as the backbone of AI production
Tracking prompts during production revealed the need for structured prompt libraries, dictionaries, and metadata.
This case study is still being iterated on as we document workflows from recent shoots.
The tracking problem
Prompts generate the images, but the production problem is knowing which prompt created which result. Early on we stored prompts in GitHub and ended up with hundreds of markdown files that were hard to search and even harder to contextualize.
The prompt library system
We treat prompts like production assets. Each prompt entry stores:
- Prompt text
- Associated images
- Model/seed/workflow details
- Director notes
Prompts are tagged by character, shot type, environment, lighting, and emotion so teams can return to known combinations quickly.
Workflow translation layers
To make existing tools usable, we introduced a dictionary that maps CFG → Prompt Strength, Sampler → Image Style Engine, Seed → Variation Lock, and other filmmaker-friendly labels. Each entry pairs a plain-language explanation, a film analogy, and the technical meaning so learning the system feels gradual.