1) Initial Prompt Draft

2) Eval Dataset

Curate a dataset of questions to test the prompt’s reachability

3) Feed through Claude

Question -> Answer

4) Feed through Grader

(Question, Answer) -> Grade Avg(Grades) signifies Prompt quality

5) Change Prompt and Repeat

Graders

3 common ways to grading:

Code - Evaluate programmatically. Define criteria into codable checks.

  • Useful for:
    • Checking output lengths
    • Verifying output does(n’t) have certain words
    • Syntax validation
    • Readability scores

Model (LLM) - Ask a model to assign a score to the output, or compare two versions.

  • Useful for:
    • Response quality
    • Quality of instruction following
    • Completeness
    • Helpfulness
    • Safety

Human - Ask a human to assign a score to the output, or compare two versions.

  • Useful for:
    • General response quality
    • Comprehensiveness
    • Depth
    • Conciseness
    • Relevance