Evaluate your prompts in prod, from human-graded to model-graded

Your prompt works in demos but silently regresses in prod? LLMs don't have standard unit tests. You install Workbench (human eval), then promptfoo (automated code-graded + model-graded eval), with deterministic assertions and LLM-as-judge for subjective quality. By the end, your prompt is gated by a CI that fails as soon as a regression shows up.

Level : intermediate · Duration : 240 min

Modules

1. 1. Why evaluate a prompt, concept + 4 canonical components
2. 2. Anthropic Workbench, first human eval (1-5 rubric)
3. 3. Code-graded, deterministic assertions + classifier metrics
4. 4. promptfoo, eval framework YAML, multi-provider, custom graders
5. 5. Model-graded, LLM-as-judge with structured rubric

View on Coeurdar