Prompting
Prompt Context Isolation
Each evaluation should receive only the scene, definitions, and metric context it needs.
What it does
Prompt handling is organized around the current scene and the specific metric being evaluated. The evaluator can include relevant character, plot, arc, engine, claim, or metric definitions without dumping the whole vault into every request.
Why it matters
Smaller, isolated prompts are easier to reason about, easier to debug, and less likely to let one evaluation concern contaminate another.
- Scene-only metrics can avoid unrelated story-object context.
- Targeted metrics can receive only the definitions they need.
- Structured JSON output keeps the response surface narrow.